Objective
Your mission
Build something small, identify issues, and apply corrections using the 3-part framework — one at a time with testing between each. Output: correction_log.md.
You need: Any AI coding tool. A text editor.
Time: 30 minutes.
Step 1: learn the 3-part correction framework (5 min)
Every correction should have three parts:
1. What's Wrong — Name the specific problem. Not "it looks bad" but "the buttons are gray when I wanted green."
2. What I Want — Describe the desired state. Not "make them look better" but "make the buttons green (#10b981) with white text."
3. What to Keep — Say what NOT to change. "Keep everything else the same." This prevents the AI from reworking parts that were already correct.
Full example:
"The tip percentage buttons are gray — I want them green (#10b981) with white text. The selected button should be a darker green (#059669). The success message disappears after 1 second — make it stay visible permanently. Keep everything else the same."
Common mistakes to avoid:
- "Fix it" — the AI doesn't know what "it" is
- "Make it better" — better how?
- Stacking 5 corrections in one message — the AI will miss some and break others
- No "keep the same" — the AI may rework everything
Step 2: build something small (5 min)
Build a small app from a spec (use one from a previous lesson, or write a new one). Don't correct the first build — you're going to practice corrections on it.
Step 3: identify 2-3 issues (5 min)
Review the build and find 2-3 specific things that are wrong.
Write in your file:
## Issues Found
1. [specific issue 1]
2. [specific issue 2]
3. [specific issue 3]
Step 4: apply corrections one at a time (10 min)
For each issue, write a correction using the 3-part framework. Apply ONE, test the result, then apply the next.
Write in your file:
## Correction 1
What's wrong: [specific problem]
What I want: [desired state]
What to keep: [what NOT to change]
Result after applying: [what happened — did it work?]
## Correction 2
[same structure]
## Correction 3
[same structure]
The 3-round rule
If you've given 3 specific corrections and the build still isn't right, stop. Either your spec wasn't clear enough (fix the spec and start a new build) or the chunk is too big (split it in half). Do not do a 4th, 5th, 6th correction round — each additional round has diminishing returns as the AI's context gets muddier. A fresh build with a clearer spec will get a better result in 5 minutes than 30 more minutes of corrections.
Step 5: evaluate (5 min)
Write in your file:
## Evaluation
Did the 3-part framework produce better results than vague corrections like "fix it"? [yes/no]
How do you know? [specific comparison — e.g., "the AI only changed what I asked, not other things"]
Did you hit the 3-round limit? [yes/no — if yes, what was unclear in the spec?]
You've found three problems with a build. How should you send the corrections?
Stacked corrections get partly missed and break other things. Send one three-part correction (wrong, want, keep), test, then send the next.