stop giving your AI the same feedback

Lenny's AI Builders - LAB 0001

turn the corrections you keep repeating into one rule you can check.

i kept giving my bots feedback. often the same feedback.

so i got fed up.

i asked Codex to find the feedback i had repeated, trace it to the source, and turn it into a workflow i could inspect.

why it matters

if the same mistakes keep coming back, the fix lives in your head, not in the work.

a rule you can read stops one of those mistakes on the next draft, so you don't have to type it again.

what you get

the prompt below gets you:

• your repeated corrections, quoted and linked to where you said them.
• lasting rules kept apart from one-off requests.
• one final check before the result reaches you.

i tried this on my job-search workflow. Codex proposed five roles: manager, researcher, writer, designer and reviewer. that is one setup for one task, not a rule for every job.

the reviewer found that all 10 evidence-map claim units repeated resume content word for word. zero were unique to the artifact. that receipt caught the repetition before the application reached me.

try it

1. pick one job where the same mistakes keep coming back.
2. add the threads that contain the work and your corrections.
3. read the proposed rules before Codex changes anything.
4. test the smallest useful version on one real task.
5. compare the result with your old corrections.
6. keep only the rules that stop a repeated mistake.

copy this prompt

add your thread links where it says [add links].

read these threads: [add links]

find the feedback i repeated. quote each correction and link it to the source. separate lasting rules from one-off requests. tell me what you could not read.

show me the smallest workflow that would stop the same mistakes:
1. the rules, who owns each step, and each handoff.
2. the final check before the result reaches me.

show me the plan before changing anything.

you'll know it worked when

you have one useful rule you can reuse, and the next draft doesn't repeat the mistake it was meant to stop.

if Codex can't read a thread, it says so instead of guessing.

limits and dates

the application was not finished or submitted, and i did not measure time saved.

more agents means more tokens and more handoffs. start with better rules and one final check. add a role only when its handoff has a clear job.

LAB 0001 went public on YouTube on 18 September 2026 (Perth time). the news in this LAB is what was teased or reported then, not a promise that it shipped.

the rest of this LAB

the news half covered an OpenAI release tease, an early Opus 5.2 report, a three-day Grok Bot build, Jev, the Persona wearable and Zuckerberg on safety and trust.

sources

• OpenAI: releases teased
• Opus 5.2: one user's report
• Grok Bot: a three-day build
• Jev: structured decisions
• Persona: a wearable assistant
• Zuckerberg: safety and trust

next step

watch LAB 0001 on YouTube - the feedback tip starts around 9:26.

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