LLenny's Podcast
← All frameworks
ProductivityAnton Osika (co-founder and CEO)

Precision Prompting for AI Builders

Never tell the AI 'it doesn't work' — state exactly what you expected and which parts do and don't.

Difficulty
Easy
Time to result
~days to results
Steps
3
Confidence
90%

A communication discipline for getting good output from AI coding tools. Because the AI cannot infer intent the way a human colleague sometimes can, you must specify precisely what you want and precisely what is and isn't working. Use it every time you iterate with an AI builder like Lovable, Cursor or Claude.

Origin

Anton framed this as the single tip he'd whisper to any first-time Lovable user, noting that explaining what you expect versus what you're getting is 'even more important with AI than with the humans.' It mirrors the strong-PM skill of being clear about the problem.

Core principles

  • 01Be specific about the outcome you want, not just the vague goal
  • 02Describe the delta: what you expected vs what actually happened
  • 03The AI does not need the 'why', but it does need exact 'what'
  • 04Errors are cheap to recover from — retry fast rather than agonise

How to run it

  1. 1

    State the exact outcome you want

    Prompt with concrete detail, e.g. 'add a button on the listing which says purchase this' and specify the resulting behaviour like 'it will pop up a modal to purchase the listing'.

    Pro tip Prompting is a skill — the more precisely you name the element and the interaction, the closer the first pass lands.

  2. 2

    Diagnose by naming the working and broken parts

    When something is off, do not say 'it doesn't work'. Say what you expected, which parts are working, and which parts are not.

    Pro tip Use the tool's chat mode to ask 'am I missing something, what should I do?' when you can't articulate the gap.

    Watch out Vague complaints are the most common and most expensive mistake users make.

  3. 3

    Iterate fast and cheaply

    If the AI misreads you, correct and retry within seconds rather than treating the miss as a failure.

    Pro tip A miscommunication that would be costly with a human engineer costs you 30 seconds here — exploit that.

In the wild

The Airbnb 'buy' button

In the live demo Anton asked for a purchase button; the AI, surprised by a buy action on an Airbnb clone, produced a 'Book now' button and booking flow instead. Rather than say it was broken, the fix was to specify the exact text change to 'Buy now'.

The correction landed instantly via visual edit, demonstrating tight, specific iteration.

Common mistakes

Saying 'it doesn't work'

This gives the AI nothing to act on. Naming the expected behaviour and the specific failing part is what unblocks it.

Assuming the AI infers intent

Humans want to understand why something matters; with AI you mostly just need to be very clear about what you're doing.

Is it for you?

Best for

Non-technical builders and PMs using AI coding tools to prototype or ship products

Not ideal for

Open-ended creative exploration where you deliberately want the model to surprise you

From the transcript

don't say it doesn't work just explain exactly what you're expecting and which parts are working and which parts are not working

20:00

explaining exactly what you expect and what you're not getting is even more important with AI than with the humans

14:30

From the episode

Building Lovable: $10M ARR in 60 days with 15 people

Anton Osika (co-founder and CEO)