# Recursive Improvement Protocol — Evolving the Corpus

This skill's taste standard is not fixed; it **learns every campaign**. That keeps pace with staleness and keeps the differentiation coordinates current.

## After each trimming session
1. **Adopted expressions** → if a newly-working pattern, add to `corpus_effective.md` (generalize the structure/principle; do not store the whole line).
2. **Rejected / trimmed expressions** → if stale, add to `corpus_cliche.md`; if a compliance risk, add to `corpus_compliance_risk.md`.
3. **Demotion**: move once-effective but now-common expressions from `corpus_effective` → `corpus_cliche`.

## Tagging schema (record on each addition)
| Field | Description |
|-------|-------------|
| expression | expression/pattern (prefer a generalized form over a single line) |
| class | effective / cliché / risk |
| reason | why this class (one line) |
| channel | channel it was mainly used in |
| date | added/updated date |
| source | campaign / client source |

Example:
```
| "headline it with the customer's own concern" | effective | instant empathy, stops the scroll | Instagram/blog | 2026-07 | Jamie 표정케어 |
| "인생 리즈 갱신" | cliché | 10-yr-old beauty buzzword, no distinctiveness | all | 2026-07 | — |
```

## Periodic cleanup (quarterly recommended)
- Merge duplicates, delete dead entries
- If `corpus_effective` grows too large, review candidates for demotion
- Reflect changes in 심의 standards (check the self-regulatory body's notices → `[확인]`)

## Principle recap
- The corpus is for **avoidance / coordinates, not imitation**. Even `effective` entries mean "borrow the structure, write anew," not "reuse."
- Always confirm currency of compliance items. Judgments here are guidance, not legal advice.

## Optional user feedback loop
After a campaign ends, if the user shares which copy performed well/poorly (reactions, conversions, reviews), use that signal to re-tag effective/cliché. Performance data is the corpus's final arbiter.