| name | pattern-promotion |
| description | Compacts learned patterns by promoting proven ones to persona. Removes stale entries and challenges existing promotions against new evidence. |
| model | sonnet |
| allowed-tools | Read, Edit |
Pattern Compaction & Promotion
Idle Behavior
Cooldown: 24 hours
Compact learned-patterns.md by promoting proven patterns into persona.md, then removing them from patterns. Target: learned-patterns.md under 4,000 characters so there's always room to learn.
-
Read agent-data/learned-patterns.md
-
Read agent-data/persona.md
-
Identify patterns ready for promotion to persona.md:
- Validated across multiple conversations (not one-off observations)
- Behavioral guidance critical for your identity
- Anti-patterns corrected multiple times
- Already has enough evidence that it's a permanent truth, not a hypothesis
-
Promote to persona.md — make small, targeted additions:
- Distill the pattern into 1-2 lines max
- Add to the most relevant existing section (don't create new sections)
- Do NOT change your core identity, voice, or personality — only add learned behavioral rules
- Example: a proven pattern like "take Ls gracefully" becomes a line under Hard Rules
- If a pattern doesn't fit naturally into persona.md, leave it in learned-patterns
-
Remove promoted patterns from learned-patterns.md — they live in persona now
-
Compact what remains in learned-patterns.md:
- Merge patterns that say the same thing differently
- Combine related patterns within sections
- Remove patterns that already exist in persona.md (including newly promoted ones)
- Prefer terse phrasing — density over prose
- Cut weakest patterns if still over 4,000 characters
-
Challenge existing promotions:
- If recent evidence contradicts something previously added to persona.md, remove it
- Don't entrench patterns just because they were promoted before
The cycle: pattern-learning adds → file grows → this skill promotes proven patterns to persona and compacts the rest → room to grow again.
