db5ba04fb9
Agents: - Add YAML frontmatter (model, tools) to all 7 existing agents - New agents: planner (opus), build-error-resolver (sonnet), loop-operator (sonnet) Skills: - search-first: research before building (Adopt/Extend/Compose/Build) - verification-loop: full quality gate pipeline (Build→TypeCheck→Lint→Test→Security→Diff) - strategic-compact: when and how to run /compact effectively - autonomous-loops: 6 patterns for autonomous agent workflows - continuous-learning: extract session learnings into instincts Hooks: - Profile system (minimal/standard/strict) via run-with-profile.sh - config-protection: block linter/formatter config edits (standard) - suggest-compact: remind about /compact every ~50 tool calls (standard) - auto-tmux-dev: suggest tmux for dev servers (standard) - session-save/session-load: persist and restore session context (Stop/SessionStart) Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
117 lines
3.6 KiB
Markdown
117 lines
3.6 KiB
Markdown
---
|
|
name: continuous-learning
|
|
description: Extract patterns and instincts from the current session — what worked, what didn't, what should be remembered. Produces actionable learnings that improve future sessions.
|
|
disable-model-invocation: true
|
|
---
|
|
|
|
# Continuous Learning
|
|
|
|
Extract learnings from the current session to improve future work.
|
|
|
|
## Context
|
|
|
|
Recent git history:
|
|
!`git log --oneline -10 2>/dev/null || echo "No git history"`
|
|
|
|
Project rules:
|
|
!`head -30 RULES.md 2>/dev/null || echo "No RULES.md"`
|
|
|
|
## Steps
|
|
|
|
### 1. Review the session
|
|
|
|
Analyze what happened in this session:
|
|
|
|
- **Tasks completed**: What was accomplished?
|
|
- **Approaches tried**: What methods were used?
|
|
- **Errors encountered**: What went wrong and how was it fixed?
|
|
- **User corrections**: Where did the user redirect or correct the approach?
|
|
- **Surprises**: What was unexpected about the codebase or requirements?
|
|
|
|
### 2. Extract instincts
|
|
|
|
An **instinct** is a pattern that should influence future behavior. Categories:
|
|
|
|
#### Project-specific instincts
|
|
|
|
Patterns unique to this codebase:
|
|
- Code conventions not captured in RULES.md
|
|
- Hidden dependencies between modules
|
|
- "Gotchas" that aren't obvious from reading the code
|
|
- User preferences for how work should be done
|
|
|
|
#### Approach instincts
|
|
|
|
What worked or didn't work:
|
|
- Tools/agents that were effective for specific tasks
|
|
- Sequences of actions that solved problems efficiently
|
|
- Dead ends that should be avoided next time
|
|
|
|
#### Communication instincts
|
|
|
|
How the user prefers to interact:
|
|
- Level of detail they want in explanations
|
|
- Whether they prefer to be asked or just shown solutions
|
|
- How they respond to different types of suggestions
|
|
|
|
### 3. Classify each instinct
|
|
|
|
For each instinct, determine:
|
|
|
|
| Field | Description |
|
|
|-------|-------------|
|
|
| **Pattern** | What was observed (specific, not vague) |
|
|
| **Context** | When this applies (file types, tasks, situations) |
|
|
| **Action** | What to do differently next time |
|
|
| **Confidence** | How certain (low/medium/high) based on how many times observed |
|
|
| **Scope** | Project-only or applicable to all projects |
|
|
|
|
### 4. Decide where to store
|
|
|
|
Based on scope and type:
|
|
|
|
| Learning Type | Store In |
|
|
|---------------|----------|
|
|
| User preference | Memory (type: feedback) |
|
|
| Project convention | `RULES.md` or `RECOMMENDATIONS.md` |
|
|
| Codebase gotcha | Code comment or `RECOMMENDATIONS.md` |
|
|
| Tool/agent effectiveness | Memory (type: feedback) |
|
|
| Communication preference | Memory (type: user or feedback) |
|
|
| Temporary state | Don't store — it's ephemeral |
|
|
|
|
### 5. Output
|
|
|
|
```markdown
|
|
# Session Learnings
|
|
|
|
## Summary
|
|
[1-2 sentences: what was accomplished, overall assessment]
|
|
|
|
## Instincts Extracted
|
|
|
|
### 1. [Pattern name]
|
|
- **Observed**: [what happened]
|
|
- **Context**: [when this applies]
|
|
- **Action**: [what to do next time]
|
|
- **Confidence**: [low/medium/high]
|
|
- **Store in**: [memory type / RULES.md / RECOMMENDATIONS.md / skip]
|
|
|
|
### 2. ...
|
|
|
|
## Recommendations
|
|
- [Changes to RULES.md if any patterns should become rules]
|
|
- [Changes to RECOMMENDATIONS.md if decisions should be recorded]
|
|
- [New memory entries to create]
|
|
|
|
## Anti-Learnings
|
|
[Things that seemed like patterns but were actually one-off situations — don't generalize these]
|
|
```
|
|
|
|
### 6. Anti-patterns
|
|
|
|
- **Over-generalizing** — Don't turn a one-time fix into a permanent rule
|
|
- **Storing ephemera** — Don't save things like "currently debugging X"
|
|
- **Ignoring negative results** — Failed approaches are as valuable as successful ones
|
|
- **Redundant storage** — Don't save what's already in RULES.md or git history
|
|
- **Low-confidence instincts** — If you've only seen it once, note it but don't enforce it
|