Continuous LearningSAFE

The most comprehensive toolkit for Claude Code -- 135 agents, 35 curated skills, 42 commands, 176+ plugins, 20 hooks, 15 rules, 7 templates, 14 MCP configs, 26 companion apps, 52 ecosystem entries, and more.

The instruction file, verbatim from the audited commit — this is the text the model reads, and the surface the audit's instruction layer examines. Quoted here so you can judge it without cloning anything.

---
name: continuous-learning
description: Auto-extract patterns from coding sessions, track corrections, and build reusable knowledge with confidence scoring
---

# Continuous Learning

## Pattern Extraction Framework

After every significant coding session, extract and categorize learnings into three buckets:

1. **Corrections** - Mistakes caught during review or by the user
2. **Successful Approaches** - Patterns that worked well and should be repeated
3. **Anti-Patterns** - Approaches that caused problems and should be avoided

## Learning Entry Format

```yaml
pattern:
  id: "LEARN-2025-0042"
  category: "error-handling"
  type: "correction"         # correction | success | anti-pattern
  confidence: 0.85           # 0.0 to 1.0
  language: "typescript"
  context: "API error responses"
  observation: "Returning raw error messages from database exceptions exposes internals"
  lesson: "Always map database errors to application-level error codes before returning"
  example:
    before: "catch (e) { res.status(500).json({ error: e.message }) }"
    after: "catch (e) { logger.error(e); res.status(500).json({ error: 'INTERNAL_ERROR' }) }"
  frequency: 3               # times this pattern has been observed
  last_seen: "2025-06-15"
```

## Confidence Scoring

| Score | Meaning | Action |
|-------|---------|--------|
| 0.95+ | Verified across multiple projects | Apply automatically |
| 0.80-0.94 | Confirmed in this codebase | Apply and mention |
| 0.60-0.79 | Observed but not fully validated | Suggest with caveat |
| 0.40-0.59 | Hypothesis based on limited data | Ask before applying |
| <0.40 | Speculative, needs validation | Document but do not apply |

Update confidence based on:
- +0.10 when pattern is confirmed correct by user
- +0.05 when pattern is observed again in a different context
- -0.15 when pattern leads to a correction
- -0.20 when pattern is explicitly rejected by user

## Session Wrap-Up Protocol

At the end of each session or before context compaction:

1. **Review changes made** - Scan diffs for patterns
2. **Identify corrections** - What was changed after initial implementation?
3. **Note successful first-attempts** - What worked without revision?
4. **Record environment details** - Framework versions, config specifics
5. **Update confidence scores** - Adjust based on session outcomes
6. **Write to knowledge base** - Append new entries to CLAUDE.md or LEARNED.md

```markdown
## Session Learnings (2025-06-15)

### Corrections Applied
- [0.85] TypeScript: Use `satisfies` instead of `as` for type narrowing with object literals
- [0.90] Next.js: Server Actions must be async functions, even for synchronous operations

### Successful Patterns
- [0.80] PostgreSQL: Partial indexes on status columns reduced query time by 60%
- [0.75] React: Extracting data fetching into Server Components eliminated 3 useEffect hooks

### Anti-Patterns Identified
- [0.70] Avoid: Nesting more than 2 levels of Suspense boundaries (causes waterfall)
- [0.65] Avoid: Using `any` to suppress TypeScript errors in catch blocks (use `unknown`)
```

## Knowledge Base Organization

Structure the knowledge base by domain:

```
knowledge/
  error-handling.md      # Error patterns across languages
  testing.md             # Test patterns and anti-patterns
  performance.md         # Optimization learnings
  api-design.md          # API design decisions
  deployment.md          # Infrastructure learnings
  project-specific.md    # Current project conventions
```

Each file follows the same entry format. Deduplicate entries with matching `observation` fields by incrementing `frequency` and updating `confidence`.

## Correction Tracking

When a user corrects code or approach:

1. Record what was originally produced
2. Record what the correction was
3. Identify the root cause (wrong assumption, missing context, outdated pattern)
4. Create or update a learning entry
5. Search for similar patterns that might need the same correction

```markdown
### Correction Log
- **Original**: Used `useEffect` to fetch data on mount
- **Correction**: Moved data fetching to Server Component
- **Root cause**: Applied client-side SPA pattern in Server Component context
- **Generalization**: In Next.js App Router, prefer server-side data fetching for initial page data
- **Confidence**: 0.90 (confirmed across 4 components)
```

## Pattern Reinforcement

Track how often patterns are applied and whether they hold:

```
Pattern: "Use zod for API input validation"
  Applied: 12 times
  Confirmed: 11 times
  Corrected: 1 time (edge case with file uploads)
  Confidence: 0.92
  Status: ESTABLISHED
```

Statuses:
- **EMERGING** (frequency < 3) - New pattern, needs validation
- **GROWING** (frequency 3-7) - Building evidence, apply with mention
- **ESTABLISHED** (frequency 8+, confidence > 0.85) - Apply automatically
- **DEPRECATED** - Once valid, now superseded by a better approach

## Integration with Memory Files

Store learnings in the project's memory file (CLAUDE.md or equivalent):

- High-confidence learnings (>0.85) go in the main instructions section
- Medium-confidence (0.60-0.84) go in a dedicated "Learnings" section
- Low-confidence (<0.60) stay in session notes until validated
- Deprecated patterns move to an archive section with reason for deprecation

Review and prune the knowledge base monthly. Remove entries that have not been referenced in 90 days and have confidence below 0.70.

SAFEgrade B · trust 89/100 Nothing in the source contradicts what it says it does. Grade A is reserved for packages that have also passed the behavioural sandbox.

What the source does

Filesystem

none-observed

Network

none-observed

Shell

none-observed

Dependencies

pinned

Secrets in source

none-found

Findings (0)

No findings outside the package's declared scope.

Gates applied: no_behavioural_pass.