Context Consolidation
Transform accumulated learnings into organized project knowledge.
Schema
Each action log entry:
json
{"timestamp": "...", "transcript": "...", "learned": "...", "context": "..."}
- •
learned: The specific command, pattern, or fix - •
context: When this applies (the trigger condition)
Process
1. Read Entries Since Last Consolidation
bash
cat .claude/action-log.jsonl | jq -s '.'
Find last "event": "consolidation" marker. Process entries after that timestamp.
2. Group by Context
Group entries with similar context values - these are related learnings.
| Context Pattern | Destination |
|---|---|
| Config/setup values | claude.md |
| Commands/procedures | skill candidate |
| Facts/constraints | claude.md |
3. Decide: Skill vs Doc
Create skill when:
- •Same context appears 2+ times
- •
learnedcontains commands or multi-step procedure - •Pattern is reusable across projects
Add to claude.md when:
- •Single occurrence
- •Project-specific config/values
- •Facts or constraints
4. Create Outputs
For claude.md entries:
markdown
## Learnings ### [Context Category] - `learned` value here
For skills:
code
skills/[name]/SKILL.md
Skill template:
yaml
--- name: [Descriptive Name] description: Use when [context]. [What it does]. --- # [Name] [learned content, expanded if multi-step]
5. Append Consolidation Marker
json
{"event": "consolidation", "timestamp": "2025-11-25T12:00:00Z", "summary": "Added X to docs, created Y skill"}
Key Principles
- •Concrete over generic: Keep actual values, actual commands
- •Context = trigger: The
contextfield IS the "use when" condition - •Skip duplicates: Don't add same learning twice
- •Minimal skills: Only create skills for repeated, multi-step patterns