4.1free~5 min

Reading what came back, like a diagnostic

Artifact: your last AI output, mapped to which prompt parts it honored or skipped

1. The output is information

A model's response isn't just an answer. It's evidence about how the model parsed your prompt. Reading it diagnostically gives you the data for your next revision.

The questions to ask:

  • Did it honor the format? If you asked for JSON and got prose, the format line lost. Move it earlier, make it terser, or add no preamble.
  • Did it stay within the bound? If you asked for three issues and got nine, the bound line lost. Make it explicit (exactly three) or add a structural cue (first issue / second issue / third issue).
  • Did it use the context? If the output ignores your stated constraints, the constraints didn't land. Move them up, repeat them at the end, or call them out explicitly.
  • Did it stay on task? If the model went off-script (started teaching when you asked for code, asked a clarifying question when you wanted output), the task verb wasn't sharp enough.

Every one of these failures maps to a part of the prompt. Identify which part, fix that part, send again.

2. The diagnostic, in practice

Same output, two ways of reading it.

textSurface read — 'it's wrong'
The output looks bad. Let me try again.

This response throws away the information the output contains.

textDiagnostic read
The output:
- Got the function signature right ✓
- Used prose instead of code-only ✗ → format constraint failed; add 'code only, no commentary'
- Included a generic error handler I didn't ask for ✗ → bound failed; add 'don't add error handling unless I asked'
- Missed the perf optimization I mentioned in context ✗ → context buried; move it nearer the task line

The second reading produces three actionable fixes for turn two. The first reading produces nothing.

3. A small habit

After every AI response, before you act on it or react to it, spend ten seconds asking:

Which part of my prompt did the model honor? Which part did it skip?

If you can answer those two questions specifically, you can write turn two specifically. If you can't answer them, turn two will be vague — and the loop won't converge.

4. Try it