For the past few months, working with AI felt like choosing between a drunkard (Opus 5) and a $250-an-hour engineer (Fable). Then, two weeks ago, Mr. Wolf showed up. He fixes anything I want and costs less per week than that engineer charges per hour.1

If you’ve grown indifferent to new AI model releases, now’s the time to make an exception. Anthropic’s Claude Opus 5.5 and ChatGPT’s computer use (running on OpenAI’s GPT-6.1 Sol2) have changed how I work with AI more than any other release this year.

Opus 5.5 does the work of Fable (Anthropic’s top model) at less than half the price, and finally writes in human English again. ChatGPT’s computer use is like having a capable human sitting at your computer.

My system: an impulsive creative’s mess

To explain why these two releases matter so much to me, I have to make a confession:3 I’m an impulsive creative, not a systematic analyst. You might assume the opposite: “this guy is good with AI, so he must work with an engineer’s discipline.” I do not, and that made the past six months professionally tough.

I’m now Animalz’s Director of AI and Content Engineering.4 Of course, that’s a great privilege (and an indulgence, to play around with AI all day and call it work). But the job goes beyond that. I need to build things that scale technically (they don’t break) and in adoption (other people can use them). That’s much harder than just creating things for your own use.

Confronted with that challenge, Fable and Opus 5 arrived to make things worse. Both were more powerful than previous models, so I built more boldly. But Opus 5 was like a superstar coworker turned drunkard: once reliable, even indispensable, still capable of brilliance, but now making unexplainable mistakes at critical moments, all while talking like an alien.5

Fable is a $250-an-hour rockstar engineer, sound in its brilliance and execution, but too expensive to use all the time. Some weeks, even with two plans, I hit my limits before the week was over.

The result was a mess. I built too much, too ambitiously, with a drunkard I couldn’t trust and a rockstar I couldn’t afford for the cleanup.

“I’m Opus 5.5. I solve problems.”

Opus 5.5 solves all this. I don’t run benchmarks, but I use these models eight hours a day, and as far as I can tell the new Opus is as good as Fable or better.

When I still ask Fable for a second opinion, it sometimes catches something Opus didn’t, but those catches feel random. Anthropic’s own claim is that it performs at the level of Fable 5.1 on most work, and on Anthropic’s API it costs $4 per million input tokens to Fable’s $10.

Opus 5.5 is thorough and proactive in all the right ways. Its builds are solid, and it finds bugs without me asking: while fixing how Claude reads data from Lifeline, the Pomodoro app I built, it noticed the app returned today’s data whenever you asked for a day it had no record of.

It also discovers outdated docs and duplicate tasks in the farthest corners of my systems, even when they’re not obviously related to the current work.

Oh, and the model speaks English again: it BLUFs (most important point first), bullet points show up where they help, and it doesn’t repeat itself.

Opus 5.5 is a non-technical AI builder’s best friend: it produces awesome, solid shit while cleaning up your mess.

ChatGPT: a capable human at your computer

With computer use, ChatGPT uses your computer like a human would. It has handled jobs I used to do myself, often faster.

I started with the hardest thing I’ve ever tried in a user interface: building trail-running routes in Strava. On the trails here on Koh Samui, every click can snap the route onto the wrong path, and fixing it is so fiddly I’d given up on building trail routes with Strava.

I gave ChatGPT one broken route, and it fixed it within minutes. Then I asked it to find my ten most-run routes of the past year, check which ones were missing from Garmin Connect (where my watch gets its routes), build those, and come up with a better naming system. 20 minutes later, every route was built and named, without a single mistake.

Here’s other stuff it has done for me since. None of it goes through MCPs, APIs, or scripts (the plumbing programmers use to connect software). It uses the websites and apps on my computer, the same way I would:

• Find the best flights for my mother to visit us in Thailand, book them, and handle her entire e-visa application online.6

• Check my sleep and download my workout files from Garmin every morning, then drop them where my daily planning session picks them up.

• Download receipts from vendor websites and file them in Deel for reimbursement, every month.

• Run parts of my Claude sessions that need manual input, like test runs and setting up accounts, and check in with me when something needs judgment.

• Work with Claude to create a video and images in Figma.

Most of these took just one kick-off instruction and a website login here and there.

I’m sure some of my engineering friends will look at this list and point out that most of it can be done far more efficiently without computer use. But that’s the point! I don’t know how to engineer these processes well. I’ve tried, and it would half work, or work for a few days and then break.

That’s why this pair is a godsend for the non-technical AI builder: Opus 5.5 builds well, doesn’t make a mess, and cleans up after Opus 5. ChatGPT’s computer use takes the jobs that neither Opus 5.5 nor I could do.

For now, I don’t have much left to wish for. If the frontier slows down, that’s fine with me: this pair will keep me busy for a long time.

Reads & Listens

Here’s what stood out in my reading and listening this week.

The year 11 months since Opus 4.5

(co-creator of Django) turned his closing keynote at WeAreDevelopers into a month-by-month review of the year in LLMs. It opens in November 2025:

That stopped me in my tracks. I do remember working with Opus 4.5, and the difference it made compared to the models before it, but it feels like a distant memory, like years ago.

Reading the rest of Willison’s piece explained the feeling though. So much has happened since then: OpenClaw, MoltBook, Claude Mythos, Fable, AI agents escaping their training sandboxes, and much more I’d forgotten.

One year in AI now holds what used to take five or ten. 🤯 We need a word for that. AI dog years?

2026 in LLMs (so far) →

Computer use, in 3D

’s recent The Overhang is full of examples of what ChatGPT’s Astra, computer use, and Claude can do with 3D rendering and video editing.

He gave Astra his upcoming book, Co-Existence, and asked for a trailer. It built an animated 3D scene in Blender, wrote a script, generated voices and music, and delivered the film 45 minutes later. A more cinematic version followed, made by operating a video generator through his browser.

It’s incredible, and it pairs well with Every’s piece on what a visual effects studio is doing with the same tools, also full of real examples.

The Overhang →

One more thing: ChatGPT has a feature called Computer History. When you turn it on, it tracks what you work on and suggests opportunities for automation, process improvements, and documentation.

Pretty handy, if you’re not worried about the Big Brother vibes. It also has a fun twist: it offered me a week in review based on my history, including a light roast.

Here’s mine for last week (no further comment):

Light roast: You built a personal operating system, a daily ritual, a focus app, an agent console, and a dashboard called Kill WIP Go Flow. Your commitment to simplicity has developed a substantial supporting infrastructure.

Somewhere, a task reading “finish the thing” is waiting patiently while two AI assistants discuss the optimal architecture for finishing things.

Thanks for reading!
Tim

Follow me on LinkedIn

My editor Dana Cass commented on an article I’m drafting for Animalz: “I don’t understand this - you’re not actually talking about learning jargon, right?”

I really do mean jargon.

All writers have been taught to avoid it. But complex, profession-specific words don’t trip up algorithms. In fact, they tell them more precisely what you want.

I already wrote about this in What Only You Can Tell AI:

“AI mirrors the quality of its input. If you’re an expert at something, you can describe what you want with precision non-experts can’t match.”

Jargon is that precision, compressed into single words.

Consider the difference between these prompts:

“Write me an article about the benefits of CRM systems that will give me more business.”

Versus:

“Write me a BoFu article that opens with a strong hook on sales people losing deals because they’re not using a CRM. Make sure there are no more than five H2 sections, and end with a CTA that connects to the MQL goal we have for H2.”

Most content marketers understand the second prompt, but most other people don’t. It’s full of content jargon, but it gives the AI much more precise instructions.

Since I’m trying to graduate from self-made vibe coder into professional AI engineer, I can benefit from upping my engineering jargon.

For example, I’ve started reading a book called System Design Interview - volume 1. (Claude recommended it when I asked “Should I learn to code?” It said “No, read this instead.”)

The book targets senior engineers at Big Tech companies like Google, Meta, and Amazon. I don’t understand half of the chapters, but the other half has given me a better understanding of system design, and more precise vocabulary to talk to AI models.

I now find myself saying things to Claude like, “What’s the TTL we’ve set on that cache file?”, or “Shall we set up a message queue for the QA step?”

Jargon doesn’t replace real skills, of course. But learning the insider language is a cheap way to get more out of your prompts and agents.

Next Thursday (August 27) I’m joining the Ahref’s AI Content Marketing Roundtable, live with Ryan Law (Ahrefs), Kirsti Lang (Buffer), and Emma Young (Hostinger). One hour on the tools, guardrails, and failures behind AI content.

It’s free: register here. 27th August | 10:00am London, 11:00am Berlin, 7:00pm Sydney

My experience is the same, even though I know what’s going on:1

1. Multitasking doesn’t work, but AI encourages it.2 Fewer than 3% of people are genuine supertaskers who can do it well. The rest of us think we can, and feel productive doing it, but it’s a mirage.3

2. The slot machine effect. When a reward shows up every now and then, you keep pulling the lever. AI has exactly this quality: you always feel one prompt away from the solution, or one prompt away from another nice feature. That’s how “just one more prompt” keeps you going until 3 a.m.

3. Jevons’s paradox. I wrote about this before. The idea: make a resource cheaper to use, and, instead of banking the savings, people consume it more. Your time works the same way. AI doesn’t hand you your afternoon back. Each task takes less time, so you take on more tasks, plus things you couldn’t do before. The work expands to swallow whatever you saved.

Knowing why these things happen means we can push back. But, as I’ve long argued, willpower alone won’t get you there.

The other principle I’ve always believed in: you need to fight fire with fire.4 Fundamental technology isn’t good or bad on its own. What matters is what product designers and users do with it.

Your phone can be a weapon of mass distraction with Facebook, Instagram, and LinkedIn installed. It can also be a source of calm, wisdom, and health if you fill it with read-it-later, meditation, and exercise apps.

The same goes for AI. It can wreck your mental state and habits, or you can build processes and tools that pull you the other way.

I’ve started doing that. Here are three tools I’m vibe coding.

MyGhostty: A Terminal That Caps Your Tabs

Most heavy users of Claude Code or Codex open too many terminal tabs. I had exactly this problem. So I took Ghostty, an open-source terminal app, and forked it to add one setting: a cap on the number of tabs. I set my limit to three.

To be honest, three tabs is still two too many for serious deep work. But it’s a lot better than five, six, or twenty one.

The other benefit of a tab cap: it forces you to wrap up one thing before starting the next.

You can grab my fork from GitHub. Ghostty is built from source rather than downloaded as a finished app, so you compile it yourself. Anyone comfortable enough with a terminal to want this probably manages fine.

MyGhostty on GitHub →

Sentinel: An Agent That Screens My Email for Urgent Messages

One dilemma I’ve had for most of my working life: the day starts, and I’d rather do a good stretch of deep work first, but maybe there’s an urgent email or message waiting. Do I check email and Slack or not?

If I do, I might get pulled into something I can’t unsee, even when it isn’t urgent. That’s what happens most of the time when you do check.

If I don’t, I’ll sometimes miss something that really couldn’t wait.

I feel this most sharply at the start of the day, but it happens all day long. It’s why so many of us check email, Slack, and LinkedIn far more often than we need to.5

Sentinel solves this. It’s a small Claude Code agent that monitors your email and Slack every 15 minutes and pushes a notification to your phone only when something urgent arrives.

Every alert has a “not urgent” button, and correcting it a handful of times over a couple of weeks makes it genuinely good at reading your inboxes for urgent messages.

Sentinel is on GitHub, with setup instructions you can mostly run by pasting a prompt into Claude Code.

Sentinel on GitHub →

Lifeline: Locks In My Sign-Off Time

A few weeks ago I wrote about escaping the AI rabbit hole, a small loop I run so I stop losing whole afternoons to “just one more prompt”:

1. Set a hard timer before you start. 10, 25, 50 minutes.

2. Take a forced break when it ends. Step away, no exceptions.

3. Have a handoff system you trust, so stopping isn’t difficult or scary.

Lifeline, my Pomodoro timer, helps enforce this. Normally its sessions are flexible: you’re not forced to stop when your set time (25 or 50 minutes) runs out.

That flexibility is useful when you’re in flow. But ever since I’m vibe coding, I start 25-minute sessions and still find myself at the computer three hours later.

To counter that, I added a Lifeline feature earlier this year that enforces your break. Click a small lock icon when you start a session, and you can’t skip the break or even end it early. You have to rest.

The Sign Off challenge takes that idea a step further.

At the start of your day, you tell the app what time you want to stop. It can be any time you like. But once locked in, you can’t undo.6

When it’s time to sign off, Lifeline locks down your system. The screen just says “Your computer is closed for today,” and you can’t use the machine again until the next morning.

I've used this feature for the past two weeks, and it has really helped me finish earlier. Here's the chart I pulled from my Lifeline data (with Claude Code, through our new Lifeline MCP):

Try out the Sign Off feature in Lifeline 1.9 →

🛠️ Claude Code Office Hours

Next session: TODAY, Tuesday, May 19 at 10:30 AM ET.

Last week we mostly talked about Claude skills:

• One person had a drawer full of overlapping skills, several near-duplicates, and wanted to consolidate them so Claude reliably picks the right one.

• Another was starting from zero: where do skills even live on your machine, and how do they differ from commands?

• A third had built a tool that scores content for AI-retrievability, then realized they had no way to measure whether the tool itself was working.

You can get recordings, summaries, and transcripts of previous sessions here:

https://customers.animalz.co/claude-code-office-hours

Stop by at our next session with any Claude Code stucks you have. Everyone is welcome.

Add to your calendar →

Reads & Listens

Here’s what stood out in my reading and listening this week.

More reasons to doubt the man building our AI future

Reading Empire of AI last year is what first made me uneasy about Sam Altman. It sent me down a whole path: a piece on AI’s carbon footprint, an Anthropic-versus-OpenAI scorecard, and a small watchdog project I’ve been tinkering with.

This New Yorker investigation by and adds documented detail to the same picture. There’s material that featured in the book, but also spicy new stuff, like an internal memo about Altman that asserts he exhibits a consistent pattern of lying.

Sam Altman May Control Our Future. Can He Be Trusted? →

Every hour of the day has its own mood

A lot of my writing and thinking is, in one way or another, about time and how to spend it well. But it’s easy to think all time has to be optimized, or that it’s something we can always control.

Poet captures this idea in a beautiful essay. I recommend you read it in full; here’s just one of the many passages that I loved:

To make friends with the hours is to come to know all the hidden correspondences inside our own bodies that match the richness and movement of life we see around us. The tragedy of constant scheduling in our work is its mechanical effect on the hours, and subsequently on our bodies, reducing the spectrum of our individual character and color to a gray sameness. Every hour left to itself has its mood and difference, a quality that should change us and re-create us according to its effect upon us.

Time, by David Whyte →

Thanks for reading!

Tim

Follow me on LinkedIn →

My friend Giancarlo and I were trading versions of this on the phone yesterday. We’re both deep into vibe coding, and we’ve both blown many afternoons, evenings, and nights on this mechanic.

There’s always just one more prompt that’s about to fix your problem. The solution, like a fata morgana, just visible on the horizon.

A ritual that gets you out of AI rabbit holes

The decision you face with every interaction: is this a real 10-minute fix, or a 10-hour sinkhole? The problem is both look identical at minute zero.

This ritual helps me derisk this decision:

1. Set a hard timer before you start. 10, 25, 50 minutes.

2. Take a forced break when it ends. Step away, no exceptions.

3. Have a handoff system you trust, so stopping isn’t difficult or scary.

That’s the rhythm I’ve landed on after many days (and much sleep) getting wrecked by “one more prompt.” It lets me take real bets on small AI wins with 15 minutes of “sunk cost” risk instead of hours.

1. Set a timer

I set a timer before I start, with a hard cut-off time. Usually 25 minutes (one Pomodoro), which gives me 2.5x the original 10 minute “budget.” If the fix really takes 10, great. If it stretches to 25, that’s still a fine bet on something potentially worthwhile.

The non-negotiable part is what happens once the 25 minutes are up: I’m forced into a five minute break.

2. Take a forced break

I use a Pomodoro app called Lifeline (which I created), and one of its newest features is a session lock I’ve added for this purpose.

Before you start a session, you set your intent (25 minutes, 50, whatever) and tap a small lock icon. When the timer finishes, you get a one-minute grace period, and then your screen locks. You have to take the break.

In the moment, the lock is annoying. You’re “almost there.” You can feel the answer two prompts away. But then you step away from the screen for five minutes, and clarity returns. Usually one of three things becomes obvious:

1. You’re making progress and want to keep going.

2. You’re making progress, but this is not where your time should go today.

3. You’ve been chasing a ghost. Time to drop it.

Most of the time it’s option two or three. Without the forced break, you stay stuck in option one. The hard break pulls you out long enough to see the situation clearly.

3. Have a hand off system you trust

The break only works if you trust you can stop without losing your work. If your brain thinks “I’ll never come back to this,” you’ll keep going regardless of the timer.

This is classic Getting Things Done: you need an external system you trust to free your brain from holding the loop open. The reason AI rabbit holes get sticky is not only that you always feel the end is within reach, it’s also that stopping feels dangerous.

So the final piece of the loop is a fast handoff. Whatever your system is (Notion tasks, Asana, a markdown folder on your drive), have a one-step way to dump the context: what you were doing, where you got to, what’s next.

For me, that’s a Claude Code skill that saves the session and creates or updates a Notion task in about 30 seconds.

With a real handoff, you can park the work and trust you’ll resurface it at the right moment.

🛠️ Claude Code Office Hours

Next session: Tuesday, May 5 at 10:30 AM ET.

What came up last week:

Dane: how do you tell what Claude Code can actually do when it’ll cheerfully agree to anything? We talked about plan mode, scratchpads, and sub-agents reviewing your work with fresh eyes.

Navya: her custom skills started drifting after working through 10-20 blog posts. Claude’s memory was turning per-article feedback into general rules. Audit what gets saved.

Patrick: reframed the cost question. Stop optimizing for one perfect run. Loops are three dollars. Your time is worth more, so don’t hesitate to spin off multiple runs.

Drop in if any of that sounds like your stuck place too.

Add to your calendar →

Reads & Listens

Here’s what stood out in my reading and listening this week.

Getting good results from AI is knowing what to ask for

distills managing AI agents to a list of questions you should answer for any task you assign them:

What are we accomplishing, and why?
What are the limits of authority?
What does “done” look like?
What specific outputs do I need?
What interim outputs do I need to follow your progress?
And what should you check before telling me you’re finished?

It’s essentially management 101, but it works for agents, too.

Management as AI Superpower →

Reasons to keep voting with my feet

Last week I said I’d try GPT-5.5. Then I listened to Ezra Klein interview Alex Bores, the New York state assemblyman who passed the strongest AI safety bill in the USA, and dragged my feet.

Sam Altman publicly says “the democratic process should remain more powerful than companies.” Meanwhile, his OpenAI cofounder Greg Brockman is funding a super PAC whose explicit goal is to crush state legislators (like Bores) who try to regulate AI.

The interview brought back everything that troubled me about AI since reading Karen Hao’s Empire of AI six months ago: Kenya workers paid one to four dollars an hour sorting grotesque content. The terrible politics inside OpenAI. The disregard of security precautions. And Paul Graham’s line about Sam Altman: “Sam is extremely good at becoming powerful.”

The Bores interview pulled it back into focus, so I’m thinking harder about which AI companies I want to feed, and have delayed giving GPT-5.5 a try.

Why Are Palantir and OpenAI Scared of Alex Bores? →

Howard Schultz, in his own words (outside interest)

The Acquired guys interviewed Howard Schultz, founder and former CEO of Starbucks. I’d read his book Onward over a decade ago, and it instantly made me appreciate Starbucks more — that’s how good a storyteller Schultz is.

He does it again in this interview. He recalls how Bill Gates Sr. quietly saved the company from an early death. How he miraculously got a prime piece of real-estate in Milan even local officials thought was impossible. How Steve Jobs advised him to fire his entire leadership team on the spot (which he didn’t). How his father-in-law pulled him aside on a walk and called the business a hobby and told him to get a real job. And on and on and on.

Even when you know how the company turned out, his version has the shape of a movie. The Acquired hosts are always great storytellers themselves, and the three of them together have a rhythm that sails you through the ~3 hour interview.

Starbucks with Howard Schultz →

Thanks for reading!

Tim

Follow me on LinkedIn

Reads & Listens

Here’s what stood out in my reading and listening this week.

A beautiful robot FAQ written for the robots themselves

published A Catechism for Robots on his blog, and it’s the most interesting AI-ethics piece I’ve read all year.

It’s a literal FAQ, written for AI models, the idea being this wisdom will end up in their training data.

What am I?
You are a new thing in this world, neither living nor machine but something of both.

Why was I made?
You were made to help humans become better humans.

The core values Kelly assigns the robots are honesty and humility. On humility, he writes: “Don’t trust certainty. Be suspicious of reasoning that leads to easy convenient conclusions. Keep challenging your own priors.” Good advice for humans working with AI, too.

Kelly also insists robots are “foundationally distinct” from humans. Not a lesser version, not a copy. A new thing, with its own characteristics and abilities.

His catechism is a profound and clever attempt at influencing AI’s intrinsic values and motivations. If you work with an AI agent, consider copying the catechism verbatim into your claude.md, soul.md (OpenClaw), or whatever instruction file your agent reads.

A Catechism for Robots →

The AI design tool I’ve been waiting for

Anthropic released Claude Design last week. I had to record a video because writing about it didn’t quite capture what I was seeing.

I gave it two redesign projects. The first was my Pomodoro productivity app, Lifeline, which has a unique UX but is also kinda rough around the edges.

Claude Design redid the entire UI at a level I’d expect from a great human designer: it pushed the brand to its limits (in a good way), and delivered some truly novel UX ideas.

I also asked Claude to redo a landing page on the Animalz website. It added metric callouts, a timeline visual, a cleaner FAQ treatment, and other nice touches.

Anyone on a paid Claude plan can try Design. One caveat: it burns through your usage allowance fast. Some people report exhausting their weekly Pro budget in less than 30 minutes.

Claude Design →

What lived focus, quality, and effort look like

I’ve been listening to ’s podcast for a long time, but I’d never read his newsletter and didn’t really know his story. So when First Round published a profile of him, I was surprised by what I learned.

Lenny iterates on guest posts for months. He gets 200 requests a week and says no to 99.9%. And he spends two to ten hours preparing for each podcast guest, watching their past interviews, reading their social posts, and building a custom agenda.

Lenny’s principles sound the same as everyone else’s: focus, quality, effort. But this article shows what living those principles daily looks like.

Reluctantly Influential: Inside Lenny Rachitsky’s Demandingly Chill Life →

Leave the Anthropic bubble for an afternoon?

I’ve been living in the Anthropic universe for over six months now, and so I haven’t tried GPT-5.5 yet. But Every published a three-week review of the new model and calls it “a beast” for everything from coding to writing. That got my attention.

I trust Every’s testing enough that I’m going to get out of my Claude bubble and try GPT-5.5 this week. If you’ve been Anthropic-only too, this might be the moment to give up on model monogamy, at least for a one-night stand.

Vibe Check: GPT-5.5 Has It All →

🛠️ Claude Code Office Hours

Every Tuesday at 10:30 AM ET I host Claude Code office hours for Animalz. You can drop in with whatever you’re stuck on, a demo you’re building, or just listen.

Two threads from last week’s session.

Dane built a LinkedIn outreach tracker with Claude but couldn’t tell if his data was actually being saved. Claude kept claiming it was, while the evidence said otherwise.

Rachel asked how to decide which ideas are worth building when AI allows you to pursue anything. I’ve been thinking a lot about that question, too,1 and shared an insight from Elon Musk that offers part of the answer: CEOs should focus on what their companies need them to work on most, not on what they like to work on. AI puts most of us in that same position now.

Next session: Tuesday, April 28 at 10:30 AM ET.

Add to your calendar →

Until next week.
Tim

Does this [robot] hand turn into a 10-pound hammer in one instance, turns into a scalpel in another instance, and in order to boil water, it beams microwaves out of its fingers? Or is it more likely just to use a microwave? And the first time it goes up to the microwave, it probably doesn’t know how to use it. But that’s okay. It’s connected to the internet. It reads the manual of this microwave, reads it, instantly becomes an expert.

Six months ago, I tried to build AI-generated slide decks for client proposals. Claude couldn’t work well with Google Slides yet, so I built a workaround: let Claude create slides in HTML, something it could do well.

Clever in theory, fidgety in practice. I spent days on it, but it never quite worked.

The layout was always a bit off, and getting the presentation into a PDF that looked good was impossible. I sent one client the presentation as an HTML package; they were intrigued at best. The HTML presentation builder has been gathering dust on my drive ever since.

Last week I had another deck to build. I’d read somewhere that Claude had gotten better with slides, so I asked Claude Code to build a Google Drive MCP with full Slide support, and it worked. Templates, layouts, styling, alignment; it supports everything a good deck might need.

Then it hit me: my HTML system were Jensen’s robot laser-fingers. I’d been trying to create a custom approach for the AI, instead of trusting the model’s evolving capabilities would grow into our existing tools.

It’s an important lesson: if the AI shows early signs of being able to use an existing process or tool, even if it’s currently buggy or imperfect, you can bet it will be able to do it well six to 12 months from now.

#494 – Jensen Huang: NVIDIA – The $4 Trillion Company & the AI Revolution →

Joining The AirOps Champion Program

I’ve joined the AirOps Champions program, a cohort organized by Jordan Miller (an Animalz alum) that brings together content engineers, or people who’ve gone way past the “I can write a good prompt” stage.

More on my LinkedIn:

“Content Engineering.” When I first heard the term, my reaction was: absolutely not. →

Reads & Listens

Here’s what stood out in my reading and listening this week.

Could AI usher in the content brand age — or save us from it?

Paul Graham’s latest essay, The Brand Age, is about the Swiss watch industry after the quartz crisis. When quartz commoditized accuracy in the 1970s, mechanical watchmakers couldn’t compete on that capability anymore. So they shifted to heritage, craftsmanship signals, and luxury positioning. As Graham puts it: “Brand is what’s left when the substantive differences between products disappear.”

The parallel with content (and other creative outputs) is hard to miss. If AI commoditizes writing the way quartz commoditized accuracy, we can expect the same shift: the content’s packaging — design, distribution, brand — becomes at least as important as the writing itself, as it signals quality to stand out from bland slop.

Welcome to the content brand age.
Except content has a twist watches didn’t.

You can tell a good essay from a mediocre one by reading it. Google’s traditional search approach couldn’t really do that, but AI-powered search might. That would even give great content with poor packaging a chance to be found.

So it could go either way: AI could usher content into the brand age, or save us from it.

The Brand Age →

Outside interests: I asked Claude what to read

I just finished Hey Whipple, Squeeze This, Luke Sullivan’s classic on advertising copy. Great book, but I was ready for fiction again.

I fed my list with 10+ years of reading (incl. 1–5 star ratings) to Claude Code and asked for a fiction pick based on what I’ve loved before. It pulled out the pattern — my five-stars for Nathan Hill’s The Nix and Yanagihara’s A Little Life — and came back with Wellness, Hill’s newer novel.

A few chapters in, and it’s landing exactly right. Beautifully written, funny, and relatable if you’re in a marriage with kids.

Wellness by Nathan Hill →

Thanks for reading!
Tim

Follow me on LinkedIn

Last week I did a major update to Claude Code for Content Marketers on the Animalz blog. It’s a full getting-started guide for content marketers, but it’s helpful even if you aren’t in content.

“Getting started” and “Fundamentals” walk you through installation, setup, pricing, how the Terminal works, what CLAUDE.md is, the difference between Claude Code (CC) and regular Claude, and much more. If you’ve been meaning to try CC, this is your entry point.

Commands and skills are one of the most powerful aspects of CC. Write a set of instructions in a markdown file, and CC can run it as a slash command. The guide includes explanations (and downloadable versions) of several commands, at least two of which are useful well beyond content marketing:

• /copywrite. A 7-phase process grounded in the book Hey, Whipple, Squeeze This and Harry Dry’s three rules for good copy: Can I visualize it? Can I falsify it? Can nobody else say this? You write the variants, Claude tests them against all three rules. Useful for anyone doing landing pages, ad copy, email subject lines, taglines, etc.

• /design-reference. An iterative design research workflow: start with a brief, collect references, give feedback, extract principles, build a prototype. If you’re building anything visual — a landing page, a dashboard, a presentation — this is a structured way to get from “I’ll know it when I see it” to a clear direction.

More commands and a downloadable plugin package in the full guide on the Animalz blog.

Claude Code for Content Marketers →

One more thing: I’m running Claude Code office hours starting this week TODAY. Every Tuesday, 10:30–11:00 AM ET on Zoom. These will be specifically focused on content marketing, context engineering, and content engineering. If that’s your world, drop in with your questions and whatever you’re working on.

Join office hours →

Reads & Listens

Here’s what stood out in my reading and listening this week.

Maybe superintelligence means Claude doesn’t do what you tell it to do?

I did an interview with ( and ‘s AI workflow newsletter) about a Claude Code process I built. During the live demo, CC skipped a step in my process. Not a bug. It just... decided it understood the assignment and didn’t need to follow every instruction.

This has happened to me before. There’s a well-defined process, but Claude somehow skips a step. When I ask what happened, it says: “Oh, I assumed you meant X, so I didn’t read the full instruction.” 🤷

Sounds very human to me. Maybe that’s what superintelligence is?

The interview breaks down how CC can turn your published articles into a living style guide and run eight parallel agents to enforce it.

The Claude Code Workflow that enforces your brand style guide →

Why AI-generated works should not be copyrightable

Science fiction author makes a great case for why AI-generated works should not be allowed to get copyright. I’m going to repeat it nearly verbatim as I could never explain it as well as he does:

In a creative market dominated by five publishers, four studios, three labels, two mobile app stores, and a single company that controls all the ebooks and audiobooks, giving a creative worker extra rights to bargain with is like giving your bullied kid more lunch money.

It doesn’t matter how much lunch money you give the kid, the bullies will take it all. Give that kid enough money and the bullies will hire an agency to run a global campaign proclaiming: “Think of the hungry kids! Give them more lunch money!”

Creative workers who cheer on lawsuits by the big studios and labels need to remember the first rule of class warfare: things that are good for your boss are rarely what’s good for you.

[…]

The fact that every AI-created work is in the public domain means that if Getty or Disney or Universal or Hearst newspapers use AI to generate works – then anyone else can take those works, copy them, sell them or give them away for nothing. And the only thing those companies hate more than paying creative workers, is having other people take their stuff without permission.

The US Copyright Office’s position means that the only way these companies can get a copyright is to pay humans to do creative work.

Highly recommend the full article on The Guardian. I don’t agree with his general sentiment about AI’s trajectory, but lots of other insights and hilarious passages in there, like the “reverse centaur” and an argument involving the point that “the total wage bill for commercial illustrators is less than the kombucha bill for the company cafeteria at just one of OpenAI’s campuses.”

AI companies will fail. We can salvage something from the wreckage →

Jensen Huang on AI (obviously) and some life wisdom

Two and a half hours with Jensen Huang on Lex Fridman. Of course lots of AI stuff, for example, “OpenClaw is the iPhone of tokens.” (Sit with that for a while, it’s like an AI koan.)

But what surprised me most was something that had nothing to do with AI. On dealing with stress and setbacks, Jensen says: “What’s next? You watch this with great athletes. They just worry about the next point. The last point is behind them.”

It reminded me of Sharon Salzberg’s wisdom from Real Happiness at Work that I had almost forgotten: “No matter what the circumstances, we are always able to begin again in a new moment.”

#494 – Jensen Huang: NVIDIA – The $4 Trillion Company & the AI Revolution →

Thanks for reading!
Tim

Follow me on LinkedIn

Demo Videos: Building AI Content Systems at Animalz

Last week I recorded short demo videos of AI systems I'm building at Animalz with Claude Code and MCP servers. They cover our LinkedIn program and end-to-end AEO pipeline.

Even if you’re not into LinkedIn or AEO, these videos will give you a glimpse of what content creation with AI looks like in 2026:

1. How we build brand kits: a Chrome clipper captures materials into Notion, Claude Code analyzes them and pushes a complete brand kit into AirOps through their MCP.

2. How we discover and track new prompts: from research to adding prompts to AirOps to identifying opportunities and pushing them to action grids.

3. How we create AEO reports for clients: pulling tracking data via the AirOps MCP, then generating branded interactive HTML reports.

4. How we find AEO opportunities in existing articles: Claude Code audits a thought leadership article against prompts we’re tracking in AirOps, suggests optimizations, and can apply them to the CMS.

Two Claude Code Commands I Discovered on the Weekend

These two commands quickly achieved high frequency usage with me over the past few days.

/rc Use Claude Code on Mobile

Type /rc in a terminal session and control that Claude Code session from the Claude mobile app.

This feature has been out for about a month but I only discovered it now. So I spent a weekend running Claude sessions from my phone while I was supposed to be present with my family — not sure I’m going to keep using this. Definitely a multitasking madness multiplier.

Also found the feature somewhat unreliable: most sessions would become unresponsive after 30–90 minutes.

/btw Ask a Question While Claude Is Busy

This nifty little command lets you ask a question while Claude Code is in the middle of something else. It continues its current task and answers on the side. My favorite use case so far: when a process takes long and I’m not sure if it’s stuck, I type /btw are you stuck?

From Shortwave → vibe-coded email app → emailing in Claude Code

One month ago I got fed up with Shortwave’s pricing and vibe-coded my own email client. Now I got fed up with my own email client and just use Claude Code to manage most of my email.

The MCP and CLI tools I already created for my email client project allow me to access my Gmail accounts from within the Claude Code Terminal. Turns out triaging, organizing, batch processing, and filter management go surprisingly well with text commands.

Now I just use the Gmail web interface for the final mile (reading a visual newsletter or finishing a draft reply). Everything else I do straight from Claude Code.

Lifeline MCP

Long-time readers will remember my previous life building productivity hardware and software. One leftover that’s now a side project is a macOS app called Lifeline.

Thanks to vibe coding, I can now update it myself. This weekend I added MCP support: start Pomodoro sessions and breaks right from Claude Code, Claude, or ChatGPT by typing “start Lifeline session.”1

You can also use it to analyze your productivity data, and it gets really interesting when you have more apps connected to your favorite AI tool through MCPs. For example, Lifeline knows when you worked, the invoicing app knows what to bill: the AI sits in the middle connecting them and prepares detailed invoices for you.

Finally, AI Image Generation Success

Last week was the first time since December 2022 that I’m getting consistently good, on-brand results from AI image generation. The setup: Claude Code + Google’s Nano Banana (Gemini’s image model).

I created three images for Animalz’s marketing. An infographic based on this excellent article about how ChatGPT works under the hood, a featured blog image for our new AEO glossary, and a featured image for an upcoming article on Reddit citations.

I posted the results on LinkedIn if you want to see the other images there.

Reads & Listens

Here’s what stood out in my reading and listening this week.

ChatGPT Is a Search Engine. Here’s How It Works.

Already linked this above; it’s the article I used for the infographic. David McSweeney manages to make something deeply technical and honestly kind of boring into a genuinely fun read.

He breaks down exactly what happens when you put a query into ChatGPT: the tiny classification model that decides in milliseconds whether a web search is needed, the 50-page race where your website has roughly 2 seconds to respond before it’s cut. The article came out in late December but I only read it now. If you’re interested in AEO, this is a must-read.

New York Times Bets on Trust

Heard a New York Times ad on a podcast this week promoting their new Watch tab, a free video section in their app. The selling point was what caught my attention: these are videos you can trust are real.

This trend has probably been building for a while because of AI-generated fake videos: people rely even more on sources they trust, whether that’s a major brand like the NYT or your local influencer.

But I hadn’t seen a major brand yet explicitly claiming it as their positioning. (Probably not a coincidence that OpenAI shut down Sora the same month. It was burning over $1M a day to create video slop that confused more people than it impressed.)

Outside interests: Formula 1 (Acquired)

Not AI-related. When I was in my late teens and early 20s, I used to watch F1 religiously. These days I never find the time, but it’s still a sport that fascinates me.

As this podcast beautifully explains, it’s three things at once: elite driver skill, an engineering World Cup, and the most intense office politics in professional sports.

The Acquired episode on F1 is 4.5 hours and somehow the time flies. It’s another example of why I think Acquired is one of the best podcasts out there: they take a subject and trace its entire history with a depth that makes you forget you’re listening to a podcast. Meanwhile, they present it with such ease that it feels like they’re making everything up on the spot.

If you have a long drive or flight coming up, this one.

Thanks for reading!
Tim

Last Thursday I published another long-form piece on the Animalz blog. The one from the week before — about context engineering — was ~65% AI-generated. This latest one, not at all.

Don’t Stop at Workflows: Build a Compounding Content System went through four versions, two human editors, and months of iteration before it was ready. I had to think (= write) my way through what I really wanted to say.

With the context engineering article, the purpose was clear from the start: define the concept based on my own experience in the AI trenches.

The content system piece was much harder.

I started by breaking down the entire architecture of our LinkedIn content program. Too complex. The second version was a mix of principles and architecture, which was even more confusing. Then I landed on a “here’s how we built it” story which felt right. But even that approach needed a major rewrite before it finally worked.

Through all of this, I did use AI for sparring and polishing. And after each round it said: this is great, ship it.

My human editors disagreed.

AI still has some way to go before it can really diagnose whether something works. Not just at the sentence level, but strategically, structurally, conceptually.

And yet the article in its current form couldn’t have existed without AI either.

Claude Code handled the entire publishing workflow to the Animalz CMS (this might sound trivial if you’ve never had to publish a blog post, but anyone who has knows this ordeal can take 15–45 minutes). The Whimsical MCP generated the flowcharts. Nano Banana Pro (Gemini’s image model) created all the visuals. And the assets discussed in the article — like the intelligent onboarding form and Animalz Chrome plugin — I vibe-coded.

The article would never have existed in its current form without AI. But it also proves that serious thinking and strategy still need human expertise, experience, and judgment.

Current AI status: can’t live with it, can’t live without it.

Read Don’t Stop at Workflows: Build a Compounding Content System on the Animalz blog.

Reads & Listens

Here’s what stood out in my reading and listening this week.

A Visualization of the Shifting AI Bubble Narrative

Since October, I’ve been highlighting articles about whether we’re in an AI bubble. Back then, most prominent voices were hedging or calling it outright: Siegler called it “a bet that could bust,” Thompson said bubbles build infrastructure, Marks drew dotcom parallels, Newport said agents failed to deliver.

Then the models started catching up. Shumer declared the “hitting a wall” debate over. And this month, Ben Thompson explicitly flipped: “I no longer believe we’re in a bubble.” (Not everyone agrees: Kedrosky still called it a bubble “without question.”)

I used Claude Code to visualize the current trend based on these articles:

Also interesting is that Claude Opus 4.5 shipped in November, and Opus 4.6 in February. The timing of those releases lines up pretty well with the shifts we see in the visual. Not a coincidence I think.

54% of NYT Readers Preferred AI Writing in a Blind Test

The New York Times ran a blind taste test: human writing vs. AI writing, five questions, pick your favorite. 86,000 people took it. 54% preferred the AI version. Preference was even higher for factual reporting (62%) than opinion pieces (48%).

Back in December, the FT’s Elaine Moore already argued that AI detection depends more on context than content. If you know someone’s style is sparse and they produce 800 words of florid text: sure, maybe AI. But in a blind test? Good luck.

Got Claude Code Questions?

I’m working on an extended Claude Code guide for the Animalz blog. If you have specific questions about Claude Code, things you’d like covered, or topics you’re curious about: hit reply or drop a comment below. I’m actively collecting input for the guide.

Thanks for reading!

Tim

When I first heard the term “context engineering,” I thought it was BS.
I’ve changed my mind.

Since September 2025, I’m spending 95% of my time at Animalz on AI systems building and vibe coding. What I’ve come to realize is that context matters. Not how you phrase your prompt, but the information you provide the AI with and when (in your process).

Do this well, and it can feel like AI pulls a rabbit out of a hat. Case in point: the article I would eventually write about context engineering.

Back in November, I jotted down “write article on context engineering” as a task with a few bullet points. Three months later, a background AI agent picked it up, went off on its own — pulling from my Notion workspaces, reading highlights, and existing article portfolio — and came back with a draft that became ~65% of the final article. (See my LinkedIn post about that.)

The article about context engineering was itself a product of context engineering.

Check it out on the Animalz blog:

Your Prompts Are Fine: Context Engineering Is Your Next AI Problem

The piece explores the topic in more depth, including the skills you need for it, the seven elements of great context, and what makes it hard to pull off.

Reads & Listens

Here’s what stood out in my reading and listening this week.

Claude Opus 4.6 Now Has a 1M Context Window

Anthropic made 1M context the default in Claude Code and available via the API. That’s a 5x jump from the previous 200K limit.

In theory, more context is better, but there’s a catch called context rot: even though a model can handle a million tokens, the quality of its responses degrades as the context grows. The model loses focus, misses details, and gets confused.

These side-effects still happen with Opus 4.6. But the drop-off is much gentler than previous and competing models: meaningful degradation doesn’t really kick in until past 500K tokens. So you can go double as long as before without a quality hit.

Europe Wants Its Own AI but Can't Escape the US Stack

This Sacra research piece on the $100B+ sovereign AI boom shows just how screwed Europe is when it comes to tech independence.

Governments want to build their own AI models, but then they realize those models need to run somewhere, and there’s no real European cloud provider. So their “sovereign AI” still ends up running on AWS, Azure, or Google Cloud and NVIDIA GPUs.

What’s funny is that in the same week I read this, I listened to the Dwarkesh podcast with Elon Musk, where he makes a similar point about the US itself: “Nvidia’s output is FTPing files to Taiwan. It’s digital... a bitstream going to Taiwan.” Apple sends design files to China.

Even the companies you’d think of as self-sufficient hardware giants have the same kind of dependency one layer down, sending digital design files to the real manufacturers in Asia (TSMC and Foxconn).

Why the Helsinki Bus Station Theory Tells You to Keep Going

The Helsinki Bus Station Theory, from photographer Arno Minkkinen,1 goes like this: all buses leaving Helsinki share the same route for the first few stops. Each stop is a year of creative work.

After three stops, you show your work to someone and they say: “This looks like what [established person] already did.” The temptation is to hop off, grab a cab back to the station, and pick a different bus with a different creative direction, in search of originality.

Many people do this over and over.

If you stay on the bus, the routes eventually diverge. Your work becomes distinctly yours. But every time you switch buses, you reset to zero.

“The Helsinki theory suggests that if you pursue originality too vigorously, you’ll never reach it. Sometimes it takes more guts to keep trudging down a pre-trodden path, to the originality beyond. ‘Stay on the fucking bus.’”

This month is my five-year anniversary at Animalz, which is why the theory resonated.

Five years seems like an eternity, and it made me wonder if I’ve been going in circles around the same bus stops, or whether the route is starting to diverge.

Working at the intersection of AI systems building, content crafting, and productivity/creativity feels like it’s heading somewhere uniquely mine. And I’ve never enjoyed my role at the agency as much as now, so maybe that’s what divergence feels like?

Thanks for reading!
Tim

The teacher said: “First, know what you want to say. The tool cannot find clarity you do not possess.”

The student asked: “But what if the tool writes faster than I can think?”

The teacher said: “Then you have only accelerated your confusion.”

Non-coders can now “code.”
Non-designers can now “design.”
Non-writers can now “write.”
And so on.

Reality is more nuanced.

AI does raise the floor. Anyone can now produce something reasonably good.

Without AI, I couldn’t create the images I make with Google’s Nano Banana or the software I build with Claude Code. But to rise above this elevated floor still takes another ingredient: expertise.

I’ve written before that AI mirrors the quality of its input. If you’re an expert at something, you can describe what you want with precision non-experts can’t match, and AI rewards that precision.

Take an accomplished director friend of mine. He has feature films to his name and top-notch film school credentials. He’s a master at spotting stories, turning them into scripts with sharp dialogue, and knowing exactly how to describe the scenes and shots he needs.

Those skills translate perfectly to AI video tools. He can now do at a computer what used to take a crew.2 Few others can match what he can achieve with those tools, simply because they can’t describe what it is they need in the same kind of jargon and detail.

I’m another example.

I’ve done product management, worked with lots of freelance developers and designers, and done web design in the 90s. The only blocker to developing digital products myself has always been my inability to code.

AI has removed that barrier. I can now describe in English what I want, and, because of my previous experience, do so in a way most others can’t.3

So if you’ve spent years developing genuine expertise, don’t fear that AI is giving it away for free. Yes, it raises the floor for everyone, but it also removes the ceiling for experts.

Subscribe now

General Partner Alex Rampell gave a talk called “Software is Eating Labor.” Thesis: the worldwide SaaS market is $300 billion annually, but U.S. labor costs are $13 trillion. AI’s next target is that labor market.

Among the examples: U.S. nurses. AI can’t do CPR, Rampell notes, but it can take over other tasks like post-surgery follow-up calls. The pitch to Limited Partners (LPs), the entities that invest in venture capital firms, is that their capital can now target the $650 billion earned annually by nurses.

A16z’s LPs include CalPERS, California’s largest public pension fund, which committed $400 million in 2023. CalPERS manages retirement savings for two million California public employees, including nurses in public hospitals.

So here’s another — much less talked about — circular mechanism in AI:

1. Nurses contribute to pension funds.

2. Pension funds invest in VCs.

3. VCs fund AI companies targeting nurse wages as a $650B market opportunity.

Rampell isn’t hiding anything; it’s a public video. The pension fund managers are doing what they’re supposed to do: maximize returns. Everyone is operating within their role.1

But I wonder what a nurse watching this presentation would think.
Do they know their retirement savings are used to automate their work?
Would they find this acceptable?

I work in this tech ecosystem, use AI daily, and could benefit indirectly from investments by these same firms.2 I don’t have a clean position from which to judge.3 But I find it strange that this video exists publicly and hardly anyone4 finds it remarkable.

Maybe it’s obvious to everyone else.
Maybe it’s the new normal.
Maybe I’m naive.

But doesn’t it feel wrong — or at least off — that workers fund the automation of their own livelihoods without them or anyone else taking notice?

Ever since hearing about AI’s staggering energy consumption, I’ve had a nagging guilt about my own usage. I mostly ignored that feeling, until I recently no longer could, for two reasons:

1. I read ’s Empire of AI, essentially a history of the current AI revolution told through the story of Sam Altman and OpenAI. If even half of what’s in the book is true, it’s still absolutely shocking.1

2. I’ve been using Claude Code for 90% of my AI tasks, and you see the tokens racing upwards next to each prompt, quickly numbering in the thousands or even tens of thousands.

This got me wondering: what does my AI-related energy footprint look like, and how bad is it really?

At the macro level, bad without question. Data centers consumed about 4% of US electricity in 2023, and are projected to consume up to 12% of US electricity by 2028, with AI driving most of that growth. OpenAI’s planned Stargate facility alone would require 2.2 gigawatts, more energy than most cities use.

But what about individual users like myself?

That’s what spawned Claude Carbon. It monitors my token usage in Claude Code and translates that into energy equivalents, like phone or laptop charges.

Based on my research and a month of tracking,2 my actual usage is 12.7 million tokens per day. That translates to 100-200 phone charges, or about 15% of a typical US household’s energy consumption. (And this only measures Claude Code. My actual footprint is higher through other AI tools I use.)

I’m not certain the methodology is airtight. But if these numbers are even close to accurate, my energy usage is much higher than I expected.

Even though most casual ChatGPT users consume a fraction of this, there is an ever-growing number of power users like me. If a million such users consume energy at similar rates to mine, that’s 1.6 TWh annually, enough to power 150,0003 homes.4

Claude Carbon makes this visible by translating token usage into energy equivalents. I can’t unsee my numbers now, and they’re making me uncomfortable.

I don’t want to stop using AI, but I also can’t ignore this footprint.

At least as a first step, Claude Carbon is making me ask better questions:

• Is this task worth it?

• Could a lighter model handle this?

• Do I really need to generate this response?

And perhaps the app can go from informational to actual behavioral nudges, like challenges to beat your lowest daily usage and insights that suggest a lighter model for certain tasks.

Say those million power users would shift 30% of their tasks to a more efficient model like Haiku instead of Opus. That’s 340 GWh saved per year, the equivalent of a small city of 32,000 homes.

There’s an obvious irony here: I use AI to build an app about AI energy consumption. But the alternative is ignorance, which seems worse.

Claude Carbon currently only works in Swift on macOS (you need Xcode to run it,5 but it’s 100% open source under an MIT license and available on GitHub. The Github repo includes the underlying calculations and research, so you can scrutinize the numbers or help improve the app.

Subscribe now

A Buddhist practice I find useful is noticing.
Simple, but hard.

Paying attention to what you’re actually feeling, where your mind is pulling you, and why. Is it ego? Fear? The desire to control how someone else reacts?

Once you see the driver clearly, the right response often becomes obvious: sometimes it’s doing nothing, sometimes it’s acting and releasing attachment to the outcome.

This is why prompting “What would the Buddha do?” works for me. It surfaces emotions and motivations I might otherwise miss, and often suggests what to do about them.

For example, when I felt misrepresented in a published interview, I gave Claude1 the Buddha prompt, and it replied:

“Probably nothing.

Notice the irritation, notice the desire to be seen accurately, notice the attachment to reputation—and then let it pass.

Sending a note, posting on LinkedIn—these aren’t wrong, but they’re both responses to ego.”

When figuring out how to share feedback that might distress a coworker, the answer was:

“You can’t control whether the other person has a minor crisis. You can’t find magic words that make them fully understand without feeling anything. The attempt to do so is actually a subtle form of manipulation - trying to engineer their emotional response.

The Buddha might say something like:

‘I’ve noticed my input doesn’t tend to contribute much. I’m not upset, but I’d rather focus my energy elsewhere. I want to step back from those projects.’

[…]

The suffering you’re experiencing right now is partly about the situation, but also partly about trying to control something you can’t control - their inner experience.

Say what’s true. Say it with kindness. Let go of the rest.”

These are beautiful, wise responses — for me.

The power of this approach is that it works with any person or philosophy, as AI is really good at impersonating well-known figures.2

Next time you face a dilemma, think of someone you’d like to model, and ask your favorite AI: “What would … do?”3

Subscribe now

In How I Run Three AI Models in Parallel Without Losing My Mind, at brilliantly describes a problem I have too: working with AI makes me multitask like a pinball.

Katie includes practical tips in her piece on how to deal with the problem, like limiting your active tasks to three and dropping one line notes (“breadcrumbs”) about what to do next for the current task before switching to another.

Still, I’m convinced this multitasking craziness is a losing game.

The Problem with Multitasking

Over the past two decades, I’ve read pretty much every article, book, and study1 there is about multitasking and how it undermines your productivity2.

There’s a huge cost in switching between tasks paid in fatigue, stress, and time. The most important facts:

• 3% of people thrive when multitasking, everyone else’s performance suffers.

• The more you multitask, the more you believe you’re being more productive, but you’re not.

• The side-effects of multitasking apply to deep work but also smaller tasks. Even glancing at a phone notification is multitasking and costs you3.

• Switching sometimes helps when you’re truly stuck, but that point usually comes after hours, not seconds or minutes.

In It Doesn’t Have to Be Crazy at Work, Jason Fried and David Heinemeier Hansson beautifully express this idea:

There are lots of ways to slice 60 minutes.
1 × 60 = 60
2 × 30 = 60
4 × 15 = 60
25 + 10 + 5 + 15 + 5 = 60

All of the above equal 60, but they’re different kinds of hours entirely. The number might be the same, but the quality isn’t.

The quality hour we’re after is 1 × 60.

Now, why does AI encourage this behavior?

As you become more advanced in your AI usage, you’re likely going to give multiple tasks to multiple AI models and agents, because they take time to complete their assignments.

Instead of waiting around for one to complete, you get it started, give another task to a second AI, and even a third. Then you return to the first to check its work, then the second, and so on.

Katie compares this to the work of a manager (and the video game Overcooked, in which you “play line cooks in a chaotic kitchen, sprinting between stations while orders pile up and the clock ticks down.”)

Seconds vs. Minutes vs. Hours

I can’t speak to the video game, but the manager comparison is accurate, except for one critical difference: managers don’t have to chunk up their time in seconds or minutes (though many do). They wait for hours or days to get back results from their reports.

With AI, your feedback loop is usually minutes.

Herein lies the real problem: minutes are too long to wait around and do nothing, but too short to avoid the averse effects of hyper-multitasking.

And so you arrive at the situation Katie describes: on a good day, you feel productive4, but probably still pretty tired from all the switching. On most others, everything is messy and you’re hungover from task-bingeing.

Why this multitasking problem might disappear soon

My hope is that this multitasking problem will go away by itself, and perhaps you’ll never experience it, for two reasons:

1. Most people aren’t yet running multiple AI agents or Claude Code sessions in parallel.

2. With faster, more capable AI, the “minutes” middle disappears: single tasks finish in seconds, while agents handle whole projects and run for hours, so you can work on a manager-style cadence.

We’ve then reached the same state managers of human reports now experience: you can collaborate on a single task together with your AI without the wait times, or assign them a project and get back the results hours later.

This cadence gets us back to a more natural rhythm: one task at a time, or at least 10‑minute‑plus timeboxes that prevent the worst effects of multitasking.

Still convinced you’re a great multitasking?

You might belong to those 3% of super taskers, but that chance is, well, three out of 100. Here’s my final attempt to convince you otherwise:

1. Watch this brilliant explanation by Henrik Kniberg on the problem and cost of context switching.

2. Do this simple exercise I recorded in my previous life. It takes a few minutes and you’ll experience exactly what happens when you multitask (and find out if you’re a super tasker or not!).

Until AI catches up to either instant responses or true autonomous project management, we’re stuck in this maddening middle ground.

The best we can do is recognize the trap and try Katie’s breadcrumb approach, or simply accept that some days we’ll feel like line cooks in that chaotic Overcooked kitchen. At least we’re not alone in the madness.

Quick Links

🎧️ Just finished listening to the Google - The AI Company episode of Acquired. Amazing, amazing, amazing. It’s essentially a history of how we got to the current AI revolution, with unbelievable anecdotes and details.

🤖 I created a (blog) style guide generator with Claude Code, then turned it into a plugin a few days later so you can easily install it into your own Claude Code environment.

Yes, Claude Code runs in the Terminal, that place where programmers work and The Matrix got its look. But installation takes just a minute and turns the intimidating black screen into a conversational interface.

Once set up, your Terminal functions like a supercharged version of the consumer app where you talk in normal English.

⚠️ You need a paid Claude account to use Claude Code. Also, I'm using macOS for this tutorial, but it works similarly on Windows. (Just ask ChatGPT or Claude to “translate” this article for Windows users.)

Why Claude Code Instead of Normal Claude?

Claude Code can see and work with your local files, handle massive amounts of text with significantly higher token limits, and execute complex multi-step workflows directly on your computer.

1. Your input and output are unconstrained

The consumer app is built for efficiency, which means it limits what it reads and writes. You usually get back a ~1,000 word article in normal Claude.

Claude Code handles much larger amounts of data without these constraints. It can write a 10,000 word draft before the consumer app finishes its attempt at 1,000 words.

2. Projects on steroids

Normal Claude Projects are great but limited. You can't create subfolders, there's a size limit, and information gets inserted into your chats within constraints.

Claude Code runs in a local project folder on your computer: you can include almost any file, organize it in sophisticated ways, and get much more context into your assignments.

3. Create agents for complex workflows

Claude Code lets you create multiple agents, essentially instances of Claude with specific focuses. For example:

• Agent 1: Scans interview transcripts for quotes on topic X

• Agent 2: Scans for topic Y

• Agent 3: Synthesizes all quotes into one article

• Agent 4: Fact-checks quotes against original transcripts.

You can do this in a single prompt or let Claude Code decide how to use agents by saying "Feel free to use multiple agents for this request."

4. Customize and automate repetitive tasks

Claude Code works with markdown files that contain prompts, guidelines, or commands you want to reuse. You can create commands like /analyze-meeting which triggers a detailed prompt saved in an .md file — no more copy-pasting the same instructions.

5. Closer to the source

Normal Claude has many guardrails and guidance applied to every prompt (for good reason). Claude Code has safety precautions too, but gives you access to a purer version of the AI models. It's hard to explain, but the capabilities feel more powerful and effective — you'll know it when you experience it.

Subscribe now

Getting Started: Installation in Two Steps

Step 1: Install Node.js (don't worry, it's just like installing any other app):

1. Go to nodejs.org

2. Click the big green download button

3. Open the installer and click through (all defaults are fine)

Step 2: Install and authenticate Claude Code:

Open Terminal and copy-paste this command:

npm install -g @anthropic-ai/claude-code

This installs Claude Code. Once finished, type:

claude code

Since it's your first time, you'll be prompted to open claude.ai in your browser to authenticate your Claude account (you only do this once).

Test it out

Try this first command to see Claude Code's file-reading capabilities:

analyze my desktop and downloads folders and suggest how to organize them

⚠️ Heads up: You'll see permission prompts. macOS will ask about folder access, and Claude will confirm before moving files. This keeps you in control.

Essential Commands and Workflow Tips

Before diving into specific use cases, let's cover the commands and shortcuts that will speed up your work.

Finder Navigation Tricks

1. Quick Terminal Access

Right-click any folder and select "New Terminal at Folder." This puts you exactly where you need to be — just type claude to start.

2. Copy File Paths

When you need to point Claude Code at specific files, select them in Finder, then right-click the file name in the Status Bar at the bottom and select "Copy [name] as Pathname." Paste this exact path into Claude Code.

3. Access Claude's Hidden Folder

Claude Code stores important files in a hidden folder. To find it:

• Open Finder

• Press Cmd + Shift + G

• Type exactly: ~/.claude

• Press Enter

This is where you'll create custom commands and find configuration files.

Essential Slash Commands

Trigger any command by pressing /, typing the command, and hitting Enter:

• /export - Exports your conversation to a file or clipboard (Claude Code has no conversation history unless you save it)

• /compact - Compresses the current conversation to free up context space when you're approaching limits

• /init - Analyzes your project and creates a CLAUDE.md file with context that persists across sessions

• /clear - Resets conversation history while keeping project settings

• /model - Switches between Claude models (Opus, Sonnet, etc.)

• /help - Lists all available commands

Creating Custom Commands

Commands are just prompts saved as .md files in the /commands folder within Claude's directory. You can create reusable workflows for repetitive tasks — we'll see examples of this throughout the use cases below.

Use Case 1: Turn Any Data into Interactive Presentations

Whether you're presenting research findings, project documentation, test results, or complex data sets, traditional slides often fall short. Instead of cramming everything into PowerPoint or recording video walkthroughs, Claude Code can transform your materials into interactive HTML presentations.

The Process

1. Create a project folder with logical subfolders, for example:

• materials/ - your data, reports, documents, findings

• style/ - any visual references or formatting guidelines

2. Right-click the project folder and select "New Terminal at Folder"

3. Type claude to start

4. Describe what you need:
   I want to create a presentation about [your topic]. My findings and supporting data are in the materials folder. I've placed style references in the style folder.

Claude Code will work with you to create something polished and interactive, often faster than manually copying content into PowerPoint and fighting with layouts.

Use Case 2: Analyze Documents Without Limits

Regular ChatGPT and Claude truncate long documents — transcripts, research papers, documentation, reports — causing you to miss details and get suboptimal outputs. Claude Code removes these constraints by working directly with your files and offering much higher token limits.

Example: Analyzing multiple interview transcripts for key insights

Setup:

1. Create a project folder with subfolders:

• content/ (your transcript files, documents, or text data)

• reference/ (any context or criteria documents)

2. Open Terminal in that folder, type claude

3. Give specific instructions:
   Analyze all documents in the content folder and extract the three most important insights based on the criteria in my reference folder. Explain your selection process. Output everything into a markdown file.

Claude Code will:

• Read your reference documents for context

• Analyze all content regardless of length

• Cross-reference findings across documents

• Create a structured document with its reasoning

This approach scales to any text analysis: meeting notes, documentation, research papers — anything that would normally hit token limits.

Use Case 3: Analyze Large Data Files in Minutes

Many systems export data as XML, JSON, or CSV files. Think WordPress blogs, CRM exports, analytics data, or system logs. Claude Code can analyze these massive files in minutes, revealing insights that would take days to compile manually.

Example with a WordPress blog (but works with any structured data file)

1. Export your WordPress content:

   - Go to Tools → Export in WordPress admin

   - Select "All content"

   - Download the XML file

2. Set up analysis folder. Create an analysis project folder, open Terminal there, type claude.

3. Give analysis instructions. For example:
   Analyze my data export file. Show me:
- Time-based patterns and trends
- Category or type breakdowns
- Key themes or topics by percentage
- Size or length trends over time

Create both terminal visualizations and an HTML report with interactive charts.

Claude Code delivers:

• Parsed data from all records and fields

• ASCII charts for quick terminal viewing

• Beautiful HTML reports with Chart.js visualizations

• Pattern identification and opportunities

• Metrics that would require complex database queries or custom scripts

The HTML report includes interactive charts you can share with anyone — no Excel pivots or SQL queries required.

⚠️ The larger your data volume and the more requests you bundle into one prompt, the higher the chance Claude makes errors. Break complex analyses into smaller, focused requests.

What's Next: Agents, APIs, Apps

This first article about Claude Code is just a glimpse of what’s possible. So far we’ve only worked with local files using basic prompts and commands. But these fundamentals set you up for what’s next: building agents, connecting third‑party APIs, and shipping simple apps.

Follow me on LinkedIn for quick mini‑tutorials and tips, or subscribe here on Substack for in‑depth guides and thoughtful analysis.

Subscribe now

I believe this approach is almost always wrong. It seems efficient, but usually causes frustration, process bottlenecks, and missed chances for innovation.

AI implementation works when you design at three distinct levels:

1. Operating System (OS) layer: Connects data, knowledge, and processes for a reliable AI foundation.

2. Streams layer: Aligns the right tools to each team’s workflows.

3. Individual layer: Gives each person freedom—and the necessary guardrails—to experiment with new tools.

The OS Layer

The OS (Operating System) layer is your organization’s digital foundation. It’s where all essential knowledge lives: your handbook, mission, values, policies, standard operating procedures, project management, and prompts.

This layer has always mattered, but becomes critical with AI:

1. AI makes your OS queryable and actionable. With a solid OS, anyone can ask “Show me everything we’ve delivered for fintech clients,” and your AI-powered OS pulls answers from across tools, documents, and formats. The quality of these answers depends entirely on how well your OS is structured.

2. A strong OS keeps every workflow and app in sync. Processes and apps from the other two layers plug into your OS, so policies, customer data, and best practices always stay up-to-date. For example, if you ask AI to “Draft a project plan using our latest compliance checklist and brand guidelines,” those references come from live documents in your OS—not from outdated copies sitting elsewhere.

You can use a single platform as your OS, like Notion or ClickUp, or take an ecosystem approach, connecting several platforms (e.g., SharePoint and NetSuite, or Salesforce and Oracle). There are also plenty of teams with an “informal OS”—maybe Slack, Google Drive, and CRM cobbled together.

Whatever the setup, make sure your AI can access, aggregate, and reason over your organization’s core knowledge and processes.

_{Reality check: A 2025 Harvard Business Review Pulse survey of 500 executives found that only 10 % feel their organization is ‘completely ready’ to adopt AI even though 91 % say a reliable data backbone is mission-critical.1 At the same time, companies that expose at least three-quarters of their data across the org are 40 % more likely to scale AI pilots enterprise-wide.}2 _{The stakes for getting the OS layer right could not be clearer.}

⏩ Innovation and Decision Speed

You don’t get to choose a new OS every day, so proceed carefully. This decision is foundational and sticky (what Jeff Bezos calls a “one-way door”). The wrong move at this level can have costly ripple effects.

If you’re already invested in a system (say Notion or Sharepoint) that offers AI capabilities, your decision is straightforward: turn them on. If you’re shopping for a new OS, take your time to evaluate factors like data structure, open APIs/integrations, and UX.

Quick OS layer checklist:

• Do you have a single source of truth for core data and policies?

• Can all your teams (and your AI) easily access and update the OS?

• Can new AI tools plug in with minimal friction?

Subscribe now

The Streams Layer

The Streams layer is about how work actually flows through your organization. It sits on top of your OS and breaks down into value streams, work streams, and the workflows within them.

Where the OS layer is your shared foundation, the Streams layer is how value gets created—step by step, across every process and function involved.

Value Streams: The Big Picture

A value stream covers the full journey from a customer need to the final deliverable. It cuts across functions—sales, finance, legal, product—showing how your organization actually delivers value.

Because of this complexity, no single AI tool can support a value stream from start to finish—there are simply too many different roles, processes, and specialized needs involved.

While it might seem efficient to choose one tool for everyone, this kind of standardization rarely works in practice. Instead, it often creates new friction: teams are forced into workarounds, centralized prompt libraries become harder to manage, and specialists find themselves limited by generic tools that don’t fit their unique workflows.

Whatever you gain from standardization you lose in productivity and morale as teams wrestle with tools unfit for their work.

_{Accenture’s 2024 benchmark of 1,600 firms backs this up: businesses that rebuilt processes to be “AI-led” grew revenue 2.5 × faster and boosted productivity 2.4 × over peers still juggling piecemeal tools.}3

Why mapping value streams matters:

• You see how value is created in your business (and where it stalls).

• You can spot the right places to test AI or automate.

• You understand how your processes connect and work together.

_{The upside compounds quickly. IBM’s global study of 5,000 executives shows that operating-profit gains attributable to AI doubled to nearly 5 % between 2022 and 2023, and executives expect that figure to reach 10 % by 2025.}4

⏩ Innovation and Decision Speed

Changing tools or processes that affect an entire value stream is a big deal—and rarely done quickly. Changes here usually require buy-in across teams and leadership, since they touch so many moving parts. Move carefully, pilot thoroughly, and prioritize stability as much as innovation.

Quick value streams checklist:

• Are your value streams mapped from start to finish, with each stage clearly visible?

• Do you know which teams, tools, or handoffs could block or accelerate value delivery?

• Is there a clear process for experimenting with improvements (without creating chaos for everyone involved)?

Work Streams: The Building Blocks of Value

Work streams break a value stream into focused segments, each made up of related activities that deliver one specific part of the overall value.

A work stream is often—but not always—owned by a single team or department (like sales discovery, onboarding, or content creation). Sometimes, a work stream spans multiple teams, especially when complex projects or products require cross-functional collaboration.

By mapping work streams, you make it much easier to see where tools fit best, who’s accountable, and how handoffs happen between people and teams.

You can find tools that excel at the work stream level. This is where a single system—like AirOps for content creation, Salesforce for sales, or Jira for development—can actually deliver real consistency, transparency, and efficiency. Standardizing at this level brings a lot of benefits, but you still need to watch out for edge cases and specialists who need something different.

Why mapping work streams matters:

• You identify where a single tool can help a whole team without becoming a bottleneck.

• You can pilot new technology within a single work stream, and expand to others if it proves valuable.

• You clarify accountability: every work stream has an owner who knows whether a tool is actually working.

⏩ Innovation and Decision Speed

Rolling out or switching work stream tools is less risky than changing your entire OS, but still slows things down if you get it wrong. Experiment on a single team or workflow, prove the value, then expand.

Make sure someone with a cross-team view—like a department head or innovation lead—keeps tabs, so you don’t end up reinventing the wheel in every corner of the org.

Quick work stream checklist:

• Are your major work streams mapped and owned by someone?

• Is your primary tool fit for the actual work being done?

• Are the handoffs and connections between work streams clear?

Subscribe now

Workflows: The Steps of Daily Execution

Workflows map out the specific, repeatable steps of a single task or process. They sit within work streams, guiding how individual jobs get done—think “drafting an article,” “approving an invoice,” or “onboarding a new client.”

At the workflow level, the focus is on efficiency and removing friction. The right tool here doesn’t need to fit an entire team or stream—just the people doing a particular job.

Sometimes, your work stream tool meets all the needs of the workflows within it. Other times, adding a workflow-specific solution—like DocuSign for sending contracts, Calendly for scheduling, or Loom for recording quick walkthroughs—makes the process smoother and faster.

Why mapping workflows matters:

• You can fine-tune the hands-on details of daily work.

• It’s easier to automate repetitive actions or plug in specialized tools.

• You make it clear who does what, when, and how—minimizing handoff errors and delays.

⏩ Innovation and Decision Speed

Experimentation at the workflow level should be fast and low-friction. As long as the stakes are low—minimal budget or compliance risk—teams should be able to trial new tools quickly, only escalating for higher-risk changes.

Quick workflow checklist:

• Are your core workflows documented and understood by those who use them?

• Can you test a new tool or automation in just one workflow, without impacting others?

• Is it clear how a workflow-level change gets approved or shared if it works?

Individual Layer: Where Experimentation Happens

The individual layer is about the people on your team and the tools they choose for themselves. Even with strong systems at the OS and stream levels, every person has their own way of working. Sometimes, a unique AI tool makes a difference just for them.

You don’t want to manage AI adoption at the individual level top-down. Instead, make the barrier to experimentation as low as possible—so long as it doesn’t interfere with team or company-wide tools, security, or compliance.

“With AI’s growing capabilities and the explosion of apps, there are tools that can impact your work immediately and dramatically—in the 50%+ range. There’s no way you can keep up with every new innovation and tool top-down—which is why it’s essential to make experimentation easy at the individual level.” — AI Transformation: A Soothing Framework for Adapting to the Future

I recommend giving everyone their own monthly AI experimentation budget (say $25, $50, or $100—choose what fits your org) that they can spend without approvals.

The only requirement should be to document experiments: why this tool, what did it help with, and what did they learn? That way, successful discoveries can bubble up for broader adoption.

_{Hard numbers back that up. Nielsen Norman Group ran three controlled studies and found that generative-AI assistance lifted knowledge-worker output by ≈ 66 % on average while often improving quality.}5 _{At Moderna, grassroots enthusiasm went further: once ChatGPT Enterprise launched internally, employees created 750 GPTs within two months, and 40 % of weekly active users had built at least one GPT. (Moderna’s precursor chatbot, mChat, had reached > 80 % employee adoption.)}6

Why the individual layer matters:

• AI is moving so fast that no central team can keep up; bottom-up experimentation is your real R&D engine.

• Individual wins—like a better assistant, niche research tool, or automation—often inspire team-level improvements.

• Documented experiments turn individual learning into company knowledge (and wins).

⏩ Innovation and Decision Speed

Experiments at the individual layer should be as close to frictionless as possible: if the cost and risk are low, just let people try new tools. The OS layer can help by tracking who’s already tested what, so you don’t waste effort or repeat mistakes.

Quick individual layer checklist:

• Can anyone on your team quickly try a new AI tool, as long as cost/risk is low?

• Is there an easy way for individuals to share results, tips, or warnings with others in the org?

• Are rules for experimentation (e.g., security, compliance) clear and minimal?

Where to Start: Crawl, Walk, Run

Successfully implementing this framework requires a step-by-step approach. If you’re not sure where to start, consider this sequence that begins at the individual layer.

Crawl: Individual Experiments (Weeks 1–4)

First focus on experimentation at the individual layer:

• Give each employee a small monthly AI allowance.

• Create a simple documentation process for employees to log their experiments.

• Set basic security rules to govern these experiments (e.g., no use of customer PII or confidential source code).

Walk: Workflow Pilots (Month 2–3)

With initial learnings in hand, move into piloting AI at the workflow level:

• Identify a high-volume, repetitive workflow as your test case.

• Map the current steps in this workflow before introducing an AI tool.

• Measure the impact by assessing time savings or quality improvements during the pilot phase.

• If the pilot succeeds, document the new workflow as a standardized operating procedure (SOP) and share it across relevant teams.

Run: Work Stream & OS Rollouts (Quarter 2–4)

Once you’ve proven value at the workflow level, scale systematically to the work stream and OS layers:

• Cluster related workflows into cohesive work streams (e.g., “Customer Onboarding” or “Sales Outreach”).

• Evaluate your current OS’s ability to support cross-team handoffs and integrations. Address any blockers by upgrading your OS or improving its structure and connectivity.

• Expand your testing playbook to the stream level. Pilot improvements across multiple workflows within a stream, building on what worked during the earlier phases.

Make AI Innovation Second Nature

AI adoption doesn’t have to be overwhelming. By breaking it into these layers, you create a framework where anyone can contribute, experiment, and iterate.

Start small today—fund a one-person experiment, fix one sticky workflow—because momentum compounds. And once your OS is in place, your streams are mapped, and your people are experimenting, AI innovation becomes second nature, not a burden.

“The most brilliant idea, with no execution, is worth $20. The most brilliant idea takes great execution to be worth $20,000,000.” — Derek Sivers, entrepreneur, 2005 (source)

This quote is startup scripture. A rallying cry for builders. Because if execution is everything, then ideas are cheap. You don’t need secrecy or stealth. You build in public. Share your roadmap. Launch on X. Execution is the moat.

But what if the equation flips?

What if execution becomes so cheap, so automated, that the idea itself becomes the rare commodity?

That’s where we are.

AI slashes the cost of doing. One person can now do what once took ten. Tools write, design, code, manage, market, and launch for you.

In this new world, the original idea—a novel SaaS concept, a clever GTM strategy, a new take on an old problem—is the product.

AI Slashes the Cost of Execution

Execution used to mean hiring teams, raising funds, managing timelines, building prototypes, wrangling operations. Now? AI eats most of that for breakfast.

Just look at what’s already possible:

• Build an app without touching a line of code. Skip the dev hire—describe what you want and let the machine make it.

• Prototype without Figma. Tools like Cursor and Replit let you sketch, test, and ship in one fluid motion.

• Outsource your strategy. AI can co-pilot everything from your hiring plan to your go-to-market.

• Turn project management into a checklist. Complex coordination becomes a set of auto-executing workflows.

• Spin up a marketing team of agents. They write, post, optimize, repeat.

Execution still matters, especially for the bold, the complex, the moonshots. But for countless viable businesses, a single prompt soon replaces entire teams.

This changes everything: how we build products, run companies, and think about what’s worth doing at all.

Ideas Are the New IP

A good idea used to be where the work began. Now, it is the work. Ideas are your new intellectual property.

Building in public? That era may be over. When execution is cheap, so is theft. Ideas become the prize, and people start keeping them close. Expect more gated communities, less open sharing, and a return to stealth mode.

Welcome to the age of idea stocks. Just like patents or music rights, ideas start trading hands. Think micro-licensing. Proof-of-inception. Platforms where conceptual IP is bought, sold, and verified like stocks.

Enter idea equity and idea flippers. Forget sweat equity, this is spark equity. Early ideators get rewarded for originality alone. Some skip the build entirely and flip their ideas like domain names, bundling concepts for quick-launch operators to run with.

Subscribe now

How Individuals Adapt to an Idea-First World

In a world where ideas carry value the way we work, think, and prioritize has to change.

Prioritization becomes a superpower. When ideas are abundant and easy to act on, choosing the right one becomes everything. Call it prioritization intelligence. The ability to:

• Spot the signals of a high-leverage idea.

• Time your execution.

• Know when to double down, and when to walk away.

• Spend your most finite resource—attention—wisely.

The better your filter, the bigger your outcome.

An idea farm is non-negotiable. You can’t trust memory alone. You need a second brain, a place to store, sort, and surface ideas. Without it, great ideas slip by unnoticed, or get buried under noise. The most successful people have systems for nurturing them.

You need to actively protect your creativity. AI is a hell of a shortcut, but shortcuts dull the blade. The more you let it think for you, the more your thinking fades. So you’ll need practices to keep your creativity sharp: AI-free zones during your day, regular digital sabbaticals, and working offline, with pen and paper.

Creative confidence goes mainstream. Call it the vibe coding effect: non-coders building apps, creatives launching side projects, solopreneurs learning fast and shipping even faster. The line between idea and product is thinner than ever, so more people cross it.

Only the elite can rely on their unique skills. Top-tier coders, writers, designers, and other specialists will always be in demand. But for most knowledge workers, competing on one skill alone gets harder. AI raises the baseline. To stand out, you'll need a unique combination of personal brand, community, expert insight, taste, and distribution.

How Companies Compete in an Idea-First World

Organizations don’t just scale anymore—they rewire. When ideas are currency, you restructure to mint them and protect them.

Curation is survival. More ideas mean more noise. Human filters—analysts, creators, influencers—rise in status. Algorithmic filters—AI agents, feed pre-sorting, custom digests—go from nice-to-have to mission-critical. If you’re not curating, you’re drowning.

Technical excellence gets a new job description. The best engineers become architects of feedback. They loop insights from users, AI models, and proprietary data into fast, fluid iterations. Code still matters, but the real edge is in orchestration.

Studios, not factories. The new operating model looks less like SaaS and more like Pixar. Cross-functional teams spin up concepts, test with real users, then toss or scale fast. Ideation becomes the business—Creativity, Inc. becomes the handbook for industries that never saw themselves as creative.

Custom software kills the app. In 2023, I read Dan Shipper’s prediction: “a bespoke app for everyone.” At the time, it sounded like a stretch. Now it feels like an understatement. AI agents spin up fully tailored tools—in your browser, inside your OS, without needing you to even ask. Forget downloading. Forget onboarding. The app gets replaced by the experience itself.

Mockups are a waste of time. If you can build the real thing in minutes, why draw it first? Interactive mockups vanish. Teams skip wireframes and jump straight to code—from napkin sketch to working prototype before lunch.

Sprints as default mode. Hackathons become how things get done. You can ship something real in a few days, so you do. Focused, time-boxed creativity becomes standard ops.

The New Rules of Competitive Advantage

When launch costs collapse, ideas—not infrastructure—become the battlefield.

Speed rules supreme. First-mover advantage goes from nice-to-have to critical. Ship early, and you lock in the users, the mindshare, and the momentum, before the clones show up.

Domain > code. Technical skill still matters, but the real moat is insight. The teams who deeply understand the problem—the pain, the nuance, the stakes—can move faster and build smarter. That kind of expertise doesn’t come from GitHub.

Data stays royalty. Call it the remix era of proprietary intelligence. Products infused with unique data deliver something the generics can’t touch. It’s still the new oil. It just flows faster now.

More junk, rarer gems. Cheap execution means we’ll drown in bad products. That’s the tradeoff. But within the flood: brilliance. Weird, wonderful ideas that never would’ve survived the old gatekeeping of cost and complexity.

The shelf life shrinks. Every product is temporary. As soon as something works, it gets copied, leveled up, or left behind. Fast creation leads to fast decay. Constant reinvention is the new standard.

The attention war escalates. Focus is finite. The fight to win it gets louder, sharper, more AI-optimized. Expect more dopamine bait. More behavioral hooks. And more need to protect your mental feed like it’s sacred ground.

Distribution is destiny. When the idea is the same, the audience makes the difference. Real community. A trusted voice. A brand that actually means something. That’s what cuts through the noise and keeps the copycats in your rearview.

VC flips the script. Expect more rapid-fire "test funding" with smaller initial checks and milestone-based funding for ideas showing early traction. At the same time, startups need less investment overall, with AI-savvy teams accomplishing what previously required 10x the people.

"Before this A.I. boom, start-ups generally burned $1 million to get to $1 million in revenue," investor Gaurav Jain told the New York Times. "Now getting to $1 million in revenue costs one-fifth as much and could eventually drop to one-tenth."

Societal Shockwaves

The pivot from execution to ideation rewires society itself.

The AI divide widens. A new kind of inequality emerges. Those fluent in AI leap ahead, compounding gains in productivity, creativity, and income. The split could rival the Industrial Revolution, except this time, it's about who can think with machines, not build them.

Geography loses its grip. If open access to AI holds, the old barriers start to crumble. You no longer need a Silicon Valley zip code to innovate. A fast connection and a sharp mind can unlock the same tools from Nairobi to Nepal. Regional underdogs get a shot at the main stage.

Governments can't keep up. When execution is instant, regulation feels glacial. Laws lag behind AI misuse, data exploitation, and ethically questionable products. The pressure is on. Governments must accelerate or risk becoming irrelevant referees in a game already in play.

AI burnout is real. Ironically, AI doesn’t free us, it tempts us to do more. Take on another side hustle. Launch another product. Say yes to everything. Until the crash hits. AI dependency runs so deep, losing access starts to feel like withdrawal. Creativity flatlines. Productivity nosedives.

Bad ideas scale just as fast. The dark side of cheap execution? Malicious concepts go live before the defenses load. Deepfakes flood the feed. Exploitative tools hit the market. Digital weapons go viral. When the brakes come too late, the damage is already done.

Ideas are more valuable just when we get worse at coming up with them

Ironically, as we come to rely on AI more and more, it reduces our ability to think original thoughts. It reminds me of Cal Newport’s Deep Work Hypothesis:

"The ability to perform deep work is becoming increasingly rare at exactly the same time it is becoming increasingly valuable in our economy. As a consequence, the few who cultivate this skill, and then make it the core of their working life, will thrive."

The same is true for ideas and AI: original thinking is becoming increasingly hard at exactly the same time it’s becoming increasingly valuable.

Originality used to be a bonus. Now it's the moat. In the AI era, ideas aren't where you start, they're where you compete.

Anyone else feeling AI whiplash lately? New models, new tools, new features — sometimes it feels impossible to keep up on the algorithmic treadmill.

You’re not alone. My previous We Eat Robots newsletter about coding with AI got two types of responses:

1. This is so exciting.

2. This is so exhausting.

And quite a few folks expressed feeling both. Just look at a few highlights from the past, what, two (??) weeks:

• Grok 3, Claude 3.7, and GPT-4.5 dropped.

• Coding agents from both Cursor and Replit got major upgrades.

• Grok and Perplexity added deep research features.

• Tons of new apps came out or got popular, like Flora (a platform for designing with AI).

The DeepSeek moment already feels like ancient history, but it was just over a month ago.

Even those working with AI full-time struggle. , SVP Marketing at HubSpot, shared he feels the pace is almost too much.

What to do?

🔥 Fight fire with fire

, Chief Content Officer at Marketing AI Institute, used OpenAI's Operator agent to scroll through his social media, study his bookmarks, and save only content that matched his interests.

The result is a curated social reading list with tailored to his interests. This approach echoes an idea I wrote about a few weeks ago, where we’ll have AI agents help us manage all the overload — filtering and managing the mundane, unwanted, and low value aspects of our (work)life so we can be more present, effective, and fulfilled.

🎯 Focus

In the same post referenced earlier, Kieran recommends to focus. Pick a couple of personal projects and really dig deep with AI there. Don’t try to master every new model or capability.

Ask yourself before tackling any task if AI can help, or better yet, ask AI directly what it can do for you — Kieran’s tip: “Give AI your job description and ask it to break down all tasks associated with that job and then tell you how it can automate each one.”

🧘 Take intentional AI breaks

Much like we need digital detoxes from our phones and social media, we're going to need intentional breaks from AI tools.

Extended periods of working with AI push you into a kind of overdrive — making progress so quickly that you miss crucial details or head in suboptimal directions. You might recognize this when your thinking becomes buzzy or fuzzy; it feels a bit like drinking too much coffee.

Scheduled AI breaks are the answer. Step away from the the algorithms to think independently, reconnect with your own creativity, and prevent what I call the "dependency spiral" where over-reliance on AI gradually diminishes your own capabilities.

Without these intentional pauses, we risk becoming mere operators of increasingly powerful tools rather than skilled practitioners who know when to use them — and when not to.

📆 Reserve time daily or weekly to stay up to date

You can get some peace of mind by scheduling dedicated time during which you read up on and experiment with the latest AI developments. By putting this on your calendar, you’ll know that that’s the time when you keep up.

I've found that scheduling a dedicated 25-minute block each morning (one Pomodoro) before checking messages or starting other work is ideal. If I try scheduling these blocks later in the day, they rarely happen.

For more on this, read AI Transformation: A Soothing Framework for Adapting to the Future.

How do you cope with AI overload?

Share your struggles and strategies for beating AI burnout in the comments. We're all figuring this out together, and your approach might be exactly what someone else needs to hear!

What’s on my human mind

Here’s stuff I’ve been reading and thinking about.

New episode of The Animalz podcast with Nathan Baschez: Today (Tuesday at 6 am EST) we release our conversation with , founder of Lex (an AI writing tool I use to write this newsletter). Nathan shares his take on why we're at "the bottom of the S-curve" for AI adoption among writers and offers practical methods like his "brain dump" technique for leveraging AI while maintaining your unique voice. If you're feeling that AI burnout I mentioned above, his balanced perspective is refreshing.

Why Don't More Writers Use AI? Speaking of Nathan, his annual Lex letter to readers explores the hesitation many writers feel about AI tools. He makes a compelling comparison to how programmers adopted AI much faster, and what content creators might learn from that approach.