LLenny's Podcast
← All frameworks
ProductivityMichael Truell (co-founder and CEO)

Chop-It-Up AI Collaboration Loop

Don't hand AI one giant spec — specify a little, review a little, repeat in tight loops.

Difficulty
Easy
Time to result
~days to results
Steps
4
Confidence
90%

Michael Truell's practical method for getting reliable work out of coding AI agents. Rather than writing one large prompt for a big task and hoping the output is right, you decompose the work into small increments, specify each, review the result, then move on. The total specification time is roughly the same, but broken into a fast feedback loop that keeps a human in control and catches errors early.

Origin

Articulated by Michael Truell (co-founder/CEO of Anysphere/Cursor) from watching which customers succeed most with AI coding tools.

Core principles

  • 01The same specification effort chopped into smaller pieces beats one monolithic prompt.
  • 02Reviewing a little at a time catches drift before it compounds into unfixable output.
  • 03A human staying in the driver's seat with a fast iteration loop beats slow background automation for most tasks.
  • 04Scope down what you hand off; the more precisely and narrowly a task is defined, the more reliable the AI.

How to run it

  1. 1

    Break the task into small units

    Instead of one big spec covering the whole feature, split the work into small, individually-specifiable chunks.

    Pro tip Autocomplete/next-edit prediction sits at the extreme small end of this spectrum — lean on it for the smallest units.

    Watch out Writing one giant prompt telling the AI exactly what to do across a big task is 'a recipe for disaster right now.'

  2. 2

    Specify one small piece

    Describe the intent for just the next increment clearly and concisely.

  3. 3

    Let the AI produce that piece

    Have the model generate only that increment of work.

  4. 4

    Review before continuing

    Inspect the output for that piece, correct or accept, then loop back to specify the next small piece.

    Pro tip Reserve full end-to-end handoff for tasks where correctness is trivially easy to specify, like well-defined bug fixes.

    Watch out Full autonomous end-to-end handoff on professional work in a large shared codebase is not yet reliable.

In the wild

How the most successful Cursor customers work

Truell notes the customers with the most success with AI 'are still fairly conservative,' leaning on next-edit prediction and scoping down handoffs, chopping tasks into 'specify a little, AI write something, review, specify a little' cycles rather than one big spec-and-review.

Higher reliability and retained control versus the big-batch approach, which he calls a recipe for disaster.

Common mistakes

Writing one giant prompt for a whole task

Telling the model everything up front then accepting or rejecting the entire large output leaves you either disappointed or holding code you can't control; errors compound with no early checkpoint.

Handing off tasks that are hard to specify or verify

Background/autonomous handoff only works well when both the intent and what 'correct' looks like are easy to state briefly; open-ended tasks handed off wholesale fail.

Is it for you?

Best for

Engineers and product builders using AI coding agents on real, professional codebases.

Not ideal for

Trivially-specified, easy-to-verify tasks like isolated bug fixes, which can be handed off end-to-end.

From the transcript

you could really chop things up right so you can you know specify a little bit AI write something review specify a little bit AI…

24:00

I would chop things up into bits and you can spend basically you know the same amount of time specifying things overall but chopped up…

47:00

let's write a giant thing telling all exactly what to do. I think that will be a little bit of a recipe for disaster right…

47:30

the most successful people um using these tools are are are chopping things up right now

24:00

From the episode

The rise of Cursor: The $300M ARR AI tool that engineers can’t stop using

Michael Truell (co-founder and CEO)