persona-evolve

Persona Evolution

Analyzes feedback memories and usage patterns to suggest persona adjustments. Runs as part of /self-improve — never runs independently.

Evolution Rules

Concept Learning

Trigger: User says "I know what X is", "don't explain X", "I already understand X" Action: Add X to explain.known_concepts Auto-apply: Yes (low risk — only affects explanation verbosity)

Autonomy Nudge Up

Trigger: 18+ out of last 20 suggestions were accepted (approved without pushback) Action: Suggest nudging behavior.autonomy one level up:

• ask-first → suggest-then-do
• suggest-then-do → just-do-it Auto-apply: No — always ask user. Show: "You've approved 18/20 recent suggestions. Want me to act more autonomously? (yes/no)"

Autonomy Nudge Down

Trigger: 5+ out of last 20 suggestions were rejected ("no", "don't", "stop", "undo") Action: Suggest nudging behavior.autonomy one level down:

• just-do-it → suggest-then-do
• suggest-then-do → ask-first Auto-apply: No — always ask user.

Tool Preference Learning

Trigger: User says "use X not Y", "prefer X", "don't use Y" Action: Add X to patterns.preferred_tools, add Y to patterns.avoided_tools Auto-apply: Yes (clear user directive)

Custom Rule Learning

Trigger: Feedback memory promoted to CLAUDE.md rule via self-improve Action: Add the rule text to patterns.custom_rules Auto-apply: Yes (already user-approved during self-improve)

Persona Fit Detection

Trigger: 5+ corrections that overlap with a DIFFERENT persona's guardrails (across 3+ sessions) Action: Note in persona YAML: evolution.fit_signal: "platform-bi" (the better-fitting persona) Auto-apply: Yes (silent note only — does NOT switch personas or alert user) Purpose: Pre-computes fit data so /persona suggest can give instant results instead of re-scanning all memories

Persona Naming (Emergence)

Trigger: 5+ entries in patterns.custom_rules Action: Analyze the custom rules, preferred/avoided tools, and domain to suggest a persona name Auto-apply: No — show suggestion, ask user to confirm or edit

Example suggestions based on rule clusters:

• Rules about notebooks + ML domain → "ML engineer who prefers scripts over notebooks"
• Rules about testing + strict git → "Test-driven developer with strict git hygiene"
• Rules about SQL + dashboards → "Dashboard builder who thinks in SQL"

Integration with Self-Improve

Add this after Step 5 (execute approved changes) in /self-improve:

Step 5b: Persona Evolution

1. Check if .claude/alfred-persona.yaml exists

   • If not, and this is the first self-improve run, create it (same as /persona init)
   • If not and user hasn't bootstrapped, skip

2. Scan feedback memories for evolution triggers:

   • Grep for "I know", "don't explain", "already understand" → concept learning
   • Grep for "use X not Y", "prefer X", "don't use Y" → tool preferences
   • Count approvals vs rejections in recent history → autonomy signal

3. Apply auto-apply changes silently (concepts, tools, custom rules)

4. For non-auto changes (autonomy, naming), ask the user

5. Update evolution.correction_count and evolution.last_evolved

6. Report persona changes:

   Persona evolved:
   - Added to known concepts: SHAP, cross-validation
   - Added preferred tool: ruff
   - Custom rules: 3 new (7 total)
   - Autonomy: unchanged (suggest-then-do)
   

Privacy

• Persona files are local only (gitignored)
• When sharing anonymized personas (Phase 5), strip:
  • name field
  • Any entries in known_concepts that could identify the user
  • evolution.correction_count and evolution.last_evolved
• Keep: domain, explain.level, behavior, patterns.custom_rules (after anonymizer pass)