Definition
You build the feature through a conversation: a small request, a careful look at the result, a correction, the next step. Each turn is cheap, and judgment is applied all the time instead of only at the start. Most real work with AI looks like this. The skill is in the quality of each review, not in the first prompt.
Example prompt
Five turns. Turn 3 catches a problem the model did not mention.
1. Create a HomeCarousel component that renders images passed as props. No timer yet.2. Good. Now advance every 5s, and clean up the timer on unmount.3. The timer keeps running when the tab is hidden. Pause it on visibilitychange.4. Add lazy loading and a fixed aspect ratio to avoid layout shift.5. Now the double flash on transition. Keep it in CSS, not JS.
What it reveals
- Debugging skill and critical reading of the output
- Ability to break work into reviewable steps
- Judgment applied in real time
Use it when
- Most day-to-day feature work
- Unclear problems where you learn by building
- Pairing with an agent on an unfamiliar codebase
Watch out for
- Endless "still broken, fix it" loops without a diagnosis
- Losing track of the overall design across many turns
- Context drift in long conversations
In an interview
Look for
- Is each step small enough to review?
- Does the candidate catch problems the model did not mention?
Ask
- At which step did you catch something the model missed?
- When would you throw this conversation away and start over?
Origin
Community practice with no single author. It mirrors iterative and incremental development, and it is how chat-based assistants are used by default.