> Disclaimer: The following views things through a Feynman-style lens, inferred from public statements. These are not Feynman's own opinions — he passed away in 1988, long before these tools existed.
Here's the scene that opens the essay: Claire, when testing products herself, always followed the "correct path" — filling in what should be filled, clicking what should be clicked, never deliberately trying to break things. Then she let Codex take over the browser, treating required fields as attack targets, and immediately dug up a bug that had been stuck for months.
The author laughs — not at her, but at all of us. We think we "understand" something when we've really just memorized its name and its shape. As Feynman's father famously said: knowing what a bird is called in every language on Earth tells you nothing about the bird.
1. Don't Be Fooled by Names
"AI agent," "computer use," "agentic workflows" — these terms sound intimidating. Strip away the labels and look at the essence: it's a program that watches pixels on a screen, moves the mouse, and types. That's it. It doesn't "understand" a webpage; it only recognizes what's clickable.
2. Give Goals, Not Instructions
Claire initially listed 25 things for the model to test. Later, she just said "go test the onboarding flow" and let the model decide how. Coverage actually got broader.
Why? Because her own assumptions were themselves blind spots. Tell it what to test, and it only tests that. Give it a goal, and it genuinely explores.
This mirrors how experiments work: set up the apparatus, pose the question to nature, then shut up. You don't tell the electron where to go — you give it a magnetic field and it decides.
A side note: Claire's husband EJ suggested having AI stop evaluating products abstractly and instead embody specific personas — PM, engineer, team lead. Testing ChatPRD this way immediately exposed a structural problem Claire knew about but had never truly *felt* from a user's perspective. Real usage beats evaluation. Ten seconds of actual use is worth a hundred pages of reports.
3. Deep Play: Maddie Reese's Projects
Maddie Reese couldn't write a line of code before. Now, using Cursor and a Raspberry Pi, she's built a pager that receives Twitter notifications, an AI receipt printer, and a "personal API."
She describes her coding level as roughly "getting by in San Diego on tourist Spanish" — she can read the gist and spot a misconnected wire, but the functions aren't hers.
Then she said something deeply Feynman-esque: routing a tweet through four services to a pager is "hardly practical. But that's not the point."
That's deep play. Feynman famously watched someone toss a plate in a cafeteria, worked out its rotation — an exercise with "no importance whatsoever" — and it eventually led to Nobel-level work. Maddie builds for fun, and real things emerge. Conversely, people who demand every project be "elegant and justified" often finish nothing.
4. Opus 5: Strong but Annoying
Claire blind-tested it against six models: it ranked first. But it's timid, apologetic, and overly dependent on human sign-off — it refused to resolve a one-line merge conflict because it was on "someone else's branch."
She asked, "Who's smarter, you or me?" Opus 5 gave a long diplomatic answer about complementary strengths. Another model, GPT-5.6 Sol, said simply: "You know what matters, I'll handle the processing — best partners."
At this level, personality matters more than capability gaps.
And how does Claire use Opus 5 now? By *avoiding conversation with it*. Running it as an async background tool produces work she's happy with; the annoying part is reading its long essays and arguing with it. Front-end design and prototypes get tossed to it, then she walks away.
That's a profound lesson in itself: the best tool is the one you don't have to deal with.
The Common Thread
The biggest advance in these AI tools isn't that they got smarter — it's that they finally let humans *stop micromanaging*.
But most people and teams haven't learned this. They've assembled all the trappings of "AI workflows" — agents, automation, prompt engineering — yet still won't let go, still micromanage, still hand over 25-item instruction lists.
That's cargo cult. However realistic your bamboo control tower, the planes won't land.
So: Do you dare give a goal and shut up? Spend an afternoon on something "useless"? Let your best tool run in the background without poking at it?
If not, no matter how powerful the tools, you're still the person waving coconut shells at the runway.