feedback becomes software

from the episode: Jev and CUA-S1 browser demos - take your AI skills with you | LAB 0004

Lenny's AI Builders - LAB 0004

give each fix a home, so the next job gets it too.

one of my thumbnails had logos that looked pasted on. i said so. the next batch looked better.

i wanted the fix to stick for the next job too.

why it matters

LAB 0004 frame at 13:01: an earlier thumbnail with logos that looked pasted on, with Lennox speaking in the corner — source: LAB 0004 episode · 13:01 - the thumbnail i was correcting

a chat can say "i'll remember." your next job may start in a new chat.

the fix needs a home: the one skill, rule or module that owns that part of the work.

what you get

the useful move is small: find the rule that made the mistake, change that rule, then check the next result.

less repeating yourself. the next run can read the changed rule.
a smaller fix. one thumbnail note does not rewrite every part of your workflow.
a check you can see. the next output either has the same problem or it doesn't.

illustration: a mistaken rule leaves a crossed-out folder and lands in a corrected one - feedback changes the one instruction that caused the error — source: LAB illustration made for these notes with ImageGen · not a screenshot from the episode

try it

name one mistake you can point to in the work.
find the instruction that controls it. if the note is only for this job, keep it here.
read the before and after. make sure the new rule fixes your note without adding things you never asked for.
save the change in that instruction, not only in the chat.
run the job again and look for the same mistake.

copy this prompt

replace [correction] with the exact thing you saw. "the logos look pasted on" is useful. "make it better" makes the next agent guess.

Lenny’s AI Builders / LAB 0004

LAB remember this: [correction]. zoom out first. find the one skill, rule, or module that owns the behavior. show the smallest before/after change, apply it there, and prove it on the next run. keep one-off feedback scoped to this job.

you'll know it worked when

the next output doesn't have the same problem.

source: LAB 0004 episode · 13:23 - the next thumbnail batch after my feedback

the test is the next output. a saved rule on its own proves nothing.

limits and dates

these were the source states shown or discussed in the recording on 20 September 2026, not a claim that every promise or demo has since shipped. LAB 0004 went public on YouTube on 23 September 2026 (Perth time).

the legal job forecast is my opinion. the thumbnail fix is an artwork result, not a business result.

the rest of this LAB

seven stories led to the practical. Jev showed a narrow browser task. CUA-S1 split form answers from the clicks. a handoff file helped the next chat pick up a long job.

Anthropic reported Claude led 26% of its measured AI research work, with none of that work fully autonomous. OpenAI announced limited access to Astra for Law.

i also talked about the Codex reset promise and a proposed way to carry agent skills between apps.

sources

next step

watch LAB 0004 on YouTube - the thumbnail fix is at 13:01.

get my public goal skill - the handoff example i use. no signup needed.

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