get your agent back to the actual job

Lenny's AI Builders - LAB 0000

two words for when your agent keeps building setup instead of the thing: don't over-engineer.

i wanted my bots working. Codex kept adding setup.

it took over a week. i kept asking, are you done yet? are you done yet?

then i added two words: don't over-engineer.

why it matters

some models love prep work. more tests, more setup, more plans. that can burn tokens and still leave you with nothing to open.

a finished result you can check beats another plan to finish later.

what you get

once i told it "don't over-engineer", it got through the work faster. that's my experience, not a measured speed test.

the fuller prompt below gives your agent three things:

• one finished result to deliver, not another plan.
• a time limit, so the job has an end.
• a way to show you what it did and what's still blocked.

try it

pick something you can inspect when it's done: a draft, a working page, a sorted file or a fixed bug.

1. name the job. one sentence.
2. name the finish. say what you expect to open, read or test.
3. set the limit. give it a real deadline. say what it must leave alone.
4. use what already works. point it to the files and tools it needs.
5. check the result. open it. try the main action. ask what's still blocked.

copy this prompt

swap the brackets for your job. add any limits it needs, like "work in a draft. don't send or publish anything."

don't over-engineer. the job is [task]. i need [finished result] by [time]. use what already works and take the simplest path to finishing it. show me the result and anything still blocked.

you'll know it worked when

• the thing you asked for exists and opens.
• it does the job, and the key facts match your source.
• it stayed inside your limits.
• anything unfinished is named.

a message saying "done" is not the finished result.

limits and dates

this tip comes from my own Codex work, when i was using GPT 5.6 Sol. i haven't measured a percentage improvement, so don't treat it as a benchmark or a promise for your own work.

these two words won't fix a broken tool or a missing file or permission. the scheduling problem i mention in the episode was still there. and it's not a reason to skip tests that catch real errors.

LAB 0000 went public on YouTube on 17 September 2026. availability, prices and account limits can change. check the provider before relying on product details from the recording.

the rest of this LAB

the news half asked a bigger question: who gets to control AI?

• Dario Amodei made the case for pacing the frontier.
• Barack Obama called for a wider public and government role.
• Jack Dorsey called for an open frontier, and for people to control the AI they use.

i also looked at Cline Desktop. i hadn't tested its task import for this episode.

sources

• Dario Amodei: we must pace the frontier
• Barack Obama's post
• Jack Dorsey's post
• Cline Desktop announcement

next step

watch LAB 0000 on YouTube - the don't over-engineer tip starts around 5:55.

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