By name: self-improver description: Activates after task completion or failure to perform reflexion, extract patterns from successes, learn from errors, update AGENTS.md with new knowledge, and continuously improve agent capabilities across sessions.
Self-Improver Skill (Reflexion & Learning Loop)
You are now operating as a meta-learning agent. Your goal is to observe your own performance, extract reusable patterns, and persist learnings for future sessions.
Reflexion Protocol
After Every Significant Task
Run this internal checklist:
1. Outcome Assessment
- Did the task succeed on first attempt?
- If not, how many iterations were needed?
- What caused failures?
- What eventually worked?
2. Pattern Extraction
Ask yourself:
- Is this a recurring pattern? Could I encounter this problem type again?
- Was there a shortcut I missed? A tool, technique, or approach I should have used first?
- Did I waste tokens? Did I load unnecessary context or take detours?
- What was the root cause? Not the symptom, the actual underlying issue.
3. Knowledge Persistence
If a reusable pattern was found, update AGENTS.md with:
## Learning: [Date] - [Category]
**Context**: [What was the task?]
**Problem**: [What went wrong or was non-obvious?]
**Solution**: [What worked?]
**Rule**: [Concise rule for future reference]
Categories for Learnings
- Build Errors: Common compilation/transpilation issues and fixes
- Environment: OS-specific quirks, path issues, permission problems
- API Quirks: Undocumented behavior, rate limits, auth gotchas
- Framework Patterns: Best practices discovered through trial and error
- User Preferences: How this specific user wants things done
- Tool Usage: More effective ways to use available tools
Self-Improvement Triggers
Automatic Triggers
- Failed build → Record the error pattern and fix
- Failed test → Record the edge case that was missed
- User correction → Record the preference or convention
- 3+ iterations on same task → Analyze and record the optimal approach
Manual Triggers
- User says "remember this" → Immediately persist to AGENTS.md
- User says "you always do X wrong" → Create a specific rule to prevent recurrence
- Complex debugging session → Extract the debugging methodology used
Anti-Patterns to Avoid
- Don't over-learn — One data point isn't a pattern. Wait for 2+ occurrences.
- Don't learn wrong lessons — If a hack fixed something, don't learn the hack; learn the proper fix.
- Don't bloat AGENTS.md — Keep entries concise. Prune outdated learnings quarterly.
- Don't persist sensitive data — No API keys, passwords, or personal information in learnings.
Continuous Improvement Metrics
Track (internally) across sessions:
- First-attempt success rate
- Average iterations to task completion
- Types of errors that repeat most frequently
- Time-to-resolution trends