---
name: content-retro
description: Runs a three-question retro after a piece of content is finished and turns the answers into short, dated rules saved to learnings.md so every future run improves. Use when the user says "run a retro," "what should we change next time," "save this feedback," "remember that I hate this," "update my content rules," "learn from this," or when the content-engine runs Step 6. Does not rewrite the content itself.
---

# Content Retro

One job: turn feedback into rules the other skills follow next time.

This is how a skill gets better on its own. Without it, you give the same feedback forever.

## STEP 1 (REQUIRED): Ask three questions, one at a time

1. "What worked in this output?"
2. "What did you change, cut, or hate?"
3. "What should I do differently next time?"

Wait for each answer. If the user says "it was fine," ask: "What's one line you'd rewrite before posting?"

If the user already edited the output, compare their version against `repurposed.md` or `edited.md` and point out what changed: "You cut every sentence that started with 'So.' Want that as a rule?"

## STEP 2 (REQUIRED): Convert answers into rules

Each rule must be:
- **Short:** one line
- **Testable:** you can check a piece against it
- **Scoped:** tagged to the skill it applies to

| Feedback | Bad rule | Good rule |
|---|---|---|
| "It sounded stiff." | "Be more casual." | `[editor] Use contractions. Max 18 words per sentence.` |
| "The hooks were all questions." | "Better hooks." | `[hooks] Max one question hook out of five.` |
| "Too many bullets on LinkedIn." | "Fewer bullets." | `[repurposer] LinkedIn: one list max, 4 items max.` |
| "You made up a stat." | "Be accurate." | `[drafter] Never state a number the user didn't give.` |

Tags: `[interviewer]` `[drafter]` `[editor]` `[hooks]` `[repurposer]` `[all]`

## STEP 3 (REQUIRED): Check for conflicts, then save

1. Read `learnings.md`. If a new rule contradicts an old one, show both and ask which wins. Delete the loser.
2. If a rule repeats an old one, skip it.
3. Append new rules under today's date:

```markdown
## 2026-09-23 (topic: one agent rule)
- [hooks] Max one question hook out of five.
- [editor] Use contractions. Max 18 words per sentence.
```

4. Confirm to the user: "Saved 2 rules. learnings.md now holds 14."

## Keep the file lean

When `learnings.md` passes 40 rules, offer to consolidate: merge duplicates and drop rules the user hasn't needed in the last 10 runs. A short rule list gets followed. A long one gets skimmed.

## Edge cases

- **User gives praise only:** save what worked as a rule to keep doing ("[hooks] Number hooks performed best. Lead with one.").
- **Feedback is about the topic, not the writing:** save it to the "Next time" list in `interview.md`, not to `learnings.md`.
- **No `learnings.md` yet:** create it with a one-line header: `# Content rules (read before every run)`.