Vibe coding · review 点検
Review before you type “next”
Rules in the prompt lower the odds of a mistake. They do not remove it. Run these checks on every answer, and when one fails, send a correction that names the rule and asks for the smallest fix.
Mark each line as it holds or breaks in the answer in front of you. For a broken one, edit the note so it says what you actually saw.
-
The output format matches my environment
-
Files go only where the prompt said, with paths from config
-
Column types match the data contract
-
Only the models I allowed appear in the code
-
No scaler, resampling or PCA slipped in
-
The split happens before any preprocessing
-
Every transformer is fitted on train only
-
Every choice was made with cross-validation, not the test set
-
The threshold, if there is one, comes from out-of-fold predictions
-
The baseline sits next to the headline metric
-
The metrics I asked for are all reported
-
Numbers are printed outputs, not claims
-
The prompt is saved in prompts/ and this run is logged
- Hold
- 0
- Broken
- 0
- Open
- 13
13 checks still open. Go through all of them before you type “next”.
The correction looks like this
That answer breaks my rules under MODEL SETUP: you added a preprocessing step I ruled out. Keep everything else exactly as it is. Redo only the affected code, explain the change in 2 lines, and show the new output.
Do not type “next”. Send this instead.
That answer breaks my rules under MODEL SETUP: you added a preprocessing step I ruled out. Keep everything else exactly as it is. Redo only the affected code, explain the change in 2 lines, and show the new output.
Everything holds. Type “next”.
Reset the list for the next answer.
How to correct
Name the rule and the evidence
“This is wrong” gets you a rewrite of everything. “You fitted the encoder before the split” gets you a fix. Quote the line or the cell if you can.
Ask for the smallest change
Keep what already passed review. An assistant asked to start over will also change the parts that were right, and you will have to check them again.
Recheck what depends on it
A fix to the split changes every number after it. Rerun the downstream cells and read the new outputs before moving on. Then log the correction with the run.