# Prompt Change Log

One row per version bump. Keep it next to your prompt library.

| Date       | Prompt ID | Version | Change                                        | Reason                                            |
|------------|-----------|---------|-----------------------------------------------|---------------------------------------------------|
| 2026-07-10 | p-0004    | 1.0     | Initial version                               | Meeting-notes-to-actions workflow                 |
| 2026-07-18 | p-0004    | 1.1     | Added "owner" extraction requirement          | Owners dropped on ~30% of notes                   |
| 2026-07-25 | p-0004    | 2.0     | Switched output to JSON array                 | Downstream script needed structured data          |
| 2026-07-27 | p-0004    | 2.1     | Trimmed system preamble to one sentence       | Reduced token cost without quality loss           |

## Rules

- Start at 1.0. Minor bump for wording/structure. Major bump for output
  shape, new required variables, or model class change.
- Never overwrite a major version — archive it.
- No log entry, no version bump.
- On rollback: add a row with "rollback to vX.Y" and the observed symptom.

## Test inputs (per prompt)

Keep 5–10 fixed inputs in a file next to the prompt. Re-run both versions
against the same inputs, on the same model, on the same day. Score each
output 0–2 on Correctness, Output shape, Length, Tone, Actionability.
Promote only if the new version scores higher overall and never zero on
Correctness or Output shape.