5 fake users broke my site before real ones could - one prompt

from the episode: they say ai video editing is solved. i tested it on my own launch video | LAB 0031

Lenny's AI Builders - LAB 0031

the exact prompt, what the 5 fake users found on my site, and the before and after.

before you launch anything, let 5 fake first-time users break it. one prompt has Claude spin up five subagents, each a different kind of first-time visitor, and they click through your app in a real browser.

i ran it on my BuilderBench page. before the fix, only 1 of the 5 fully got what they came for. after, all 5 did, and the blank screen on slow 3G went from 5.35 seconds to 3.3.

why it matters

you can't see your own app like a stranger does. you know where every button is. a first-time visitor doesn't, and they leave quietly.

this prompt gives you five strangers on demand: the impatient one, the skeptic, the mobile user, the non-native English speaker and the breaker. it ranks what they hit by how many of them hit it, so you fix the big stuff first.

what you get

one prompt you can copy. paste it into Claude Code inside your project.
screenshots of every snag. each fake user screenshots every moment they get confused, stuck or annoyed.
proof it's fixed. it fixes everything in one pass, then reruns all 5 users.

the real example: my result

my agent ran the prompt for me on 28 October 2026, in Claude Code (Opus 5.5), on a local copy of my BuilderBench page on its own branch. nothing went live until i'd looked at it.

before: 1 of 5 fully succeeded. the other four only got part of the way.
found: 33 issues, ranked by how many users hit each one.
the top two hit all 5 users: the page told you who won one test, not which AI is best for your job, and the way to the full result was weak (a "see why" link that only scrolled).
mobile on slow 3G: blank screen went from 5.35 s to 3.3 s, and the landing page went from about 4.1 MB to 0.4 MB.
after: all 5 of 5 succeeded on the rerun.

since filming, those fixes went live. try the new BuilderBench page yourself.

try it

open Claude Code inside your project.
paste the prompt below, and add your app's address and what "first success" means for a new visitor.
let it run the five users.
read the ranked list, then let it fix and rerun.
look at the branch yourself before you ship anything.

copy this prompt

paste this into Claude Code inside your project:

Lenny’s AI Builders / LAB 0031

Act as my launch QA lead. Spin up 5 subagents, each a different first-time user:
1. The impatient one: skims, clicks the first button, quits at the first error
2. The skeptic: reads pricing, terms and privacy before signing up
3. The mobile user: does everything on a 375px screen on slow 3G
4. The non-native English speaker: only follows what the UI literally says
5. The breaker: pastes junk into forms, hits back mid-flow, double-clicks everything
Each one opens the app in a real browser, goes from landing page to first success, and screenshots every moment they get confused, stuck or annoyed.
Merge everything into one list ranked by how many users hit it. Fix it all in one pass, then rerun the 5 users to prove it's fixed.
Commit on that branch only. Don't push, deploy, or open a PR. I'll review first.

this is the prompt from the episode (it starts at 11:46). the idea came from a post by RoundtableSpace.

you'll know it worked when

you get one ranked list with a count of how many users hit each snag, before and after screenshots for each user, and a rerun where all 5 reach first success.

a bad result: it pushes or deploys anything, skips the rerun, or "fixes" things it never showed you. stop it and run it again.

what to do next

launching soon: run it on your sign-up and first-use pages before anyone real sees them.
already live: run it on the page you most want people to finish, like pricing or sign-up.
no app yet: save the prompt. it works on any site you can run locally.

limits and dates

one run. one page (BuilderBench), 28 October 2026, run by my planning agent. not a full audit.
simulated users, not real visitors. they find a lot, but they're not your customers.
not everything got fixed. a few items were left alone on purpose, like my privacy page and house style.
review before you ship. keep the fixes on a branch and check them yourself.

the rest of this LAB

Opus 5.5 vs Terra vs Astra, same raw footage. one prompt: make me a Short i'd post. i picked Opus blind. Opus took 11 min 13 s and both judges gave it 8/10. Astra took 3:21 (8 and 7), Terra 3:22 (6 and 5). Opus cost $2.52 (API-equivalent); at API list prices Astra works out to about $5.77 and Terra about $0.33. my take: taste wins.
Artificial Analysis index v5. Opus 5.5 leads at 58, then GPT-6 Astra at 57 and GPT-6.1 Sol at 55. 60% of the new version is private tests. the catch: "first model to stay #1 for a month" is a creator's claim i couldn't verify. my take: usability still beats benchmarks.
DeepSeek v4.1 Flash is cheap. 15 cents per million tokens in and 60 cents out, off-peak. the catch: cheap cuts both ways - it's cheap for attackers too.
ChatGPT plugs into iCloud. it can now connect to your iCloud mail and calendar. the catch: there's also a third-party iCloud app in ChatGPT that isn't OpenAI's. my take: connectors are the new moat.
Grok Bot got a big update. 1Password, new Apple access, and it can make videos. my take: video is the next frontier.
the ambitious skill. inspired by Felix Rieseberg's post. in my blind test it made Claude's work about 17% better. the catch: it uses more tokens.
also in this episode: a prompt that finds the promises you forgot to keep in two weeks of messages (sample data, 49 seconds).

sources

next step

watch LAB 0031 on YouTube - the prompt starts at 11:46.

ran it on your app? tell me under the video how many of the 5 made it before the fix.

want the next one? get the next field notes. optional - these notes stay free to read.