Field Notes
Hard-won lessons from building AI products.
Short pieces from building AI products in production. Every note is a perspective on how the work is changing or a tactic you can apply to a live product.
Field Notes is open to other builders. If you are building with AI and have a note worth sharing, reach out to contribute. Contributions are published under the Contributions section of our terms: you keep your rights, and we publish with credit to you.
18 notes
- Aug 162026
Why We Stopped Letting AI Pick the Winner
For a few weeks we let Claude generate creative options and let a judge agent pick the best one. The picks were competent, and we rejected nearly every one. We also cover what the judge still does for us, and why we took the final pick back ourselves.
Perspective5 min read - Jul 262026
Why We Run a Research Brief Before Every Migration
We wanted FuelTheFam's app on Cloudflare and its database with it, so before either migration we ran a research brief. The first told us to upgrade the app before moving it, and the second told us not to move the database at all. We spent the time we saved hardening the stack we already had.
Tactics5 min read - Jul 182026
How We Spent a Week's Tokens on Fable 5's Great Idea
Fable 5 offered to test our new Claude Skills in a way no earlier model had, and we let the test run through most of a week's model allowance. We cover what it suggested, why the bill was so far off the estimate, and the habits we keep from it.
Perspective6 min read - Jul 82026
How We Keep Our AI Drafts From Sounding Like AI
Every AI writing tool produces the same handful of tics, and we were deleting them by hand from every draft. So we wrote the tics down once, as a plain set of rules any tool can follow, and put the file where anyone can take it.
Tactics5 min read - Jun 182026
How We Used Loops Instead of One-Off Prompts
Using a model one prompt at a time makes you the loop. It works better when you hand the loop to the system, with an objective, a metric to score against, and a boundary for when to stop. These are the loops we ran to build this site.
Tactics4 min read - Jun 172026
The AI-Native PM's Ownership Map
The AI-native PM owns the product's behavior, not its model, and much of that ownership stays unassigned until something breaks and you find the gap. This is the working map of what you own across Shape, Ship, and Track, the move you make to deliver each one, and how you know it is done.
Tactics6 min read - Jun 162026
How We Used Agent Fleets to Build This Site
The picture of building with AI is one person and one chat window. We built this site the other way, with many workers running at once and the two of us setting the brief and judging what they returned. That is what the work looks like once you are building at any real size.
Perspective5 min read - Jun 92026
Access Control Is a Product Decision, Not a Security One
When an AI product pauses to ask "Send this?" or fires the action silently, that single screen is the whole product. What your AI is allowed to do, on whose behalf, and with which data is the trust contract the user feels, and the people who treat it as a backend chore are the ones who end up in the headlines.
Tactics8 min read - Jun 42026
How We Caught AI Mistakes With a Second Agent
When workers run in parallel, each hands back something that looks finished, and finished-looking is not the same as correct. So we sent every draft on this site to a second worker whose only job was to take it apart first.
Tactics5 min read - May 282026
Why We Added a Dark Mode to This Site
We did not add dark mode to look modern. We added it because of who reads this and when, and starting from the reader's real situation instead of from fashion is also why we built it the way we did.
Tactics5 min read - May 202026
Why We Told the AI to Disagree With Us
The most useful instruction we gave the assistant building this site was not a task. It was a standing permission to disagree, because a collaborator who only ever says yes tells you nothing.
Tactics5 min read - May 142026
The PM's Toolkit for AI Behavior Design
Two teams point the same base model at the same task and ship products that behave nothing alike. The difference is a set of levers, and the work is turning them with intent.
Tactics8 min read - May 92026
What We Built for This Site and Then Deleted
Some of the best decisions here are things you will never see, because we built them and then cut them. Deciding what to throw away is the same judgment as deciding what to make, and it is harder when the making was cheap.
Perspective5 min read - Apr 302026
How We Used Preview Labs to Design This Site
We built almost nothing here by going straight to a single answer. For each piece we laid out a set of options in a preview lab first, because when a tool can produce ten finished-looking versions in a minute, the work is no longer making one, it is choosing the right one.
Perspective5 min read - Apr 272026
Strategy Used to Be a Document. Now It's a Discipline.
In late 2023, OpenAI shipped one ChatGPT update that turned a whole cohort of startups' only differentiator into a checkbox. This is what strategy has to become when the ground can move on the platform's schedule, not yours.
Perspective7 min read - Apr 182026
Why We Replaced Our Clever Navigation With a Plain One
Twice we built a clever navigation and twice we cut it for a layout readers already knew. Every new mental model is a tax on attention, and on an education site that attention belongs to the lesson.
Tactics5 min read - Mar 212026
Taste Is a Muscle, Not a Gift
For twenty years, making one good option was the hard part. Now a tool produces ten in a minute and the whole job is choosing, which means taste, the muscle that does the choosing, is the one you most need to build.
Perspective8 min read - Mar 42026
The Missing Layer
AI put building in everyone's hands. Whether you ship or stall still comes down to a thin layer of knowledge, and it is the most learnable gap there is.
Perspective7 min read