compound-learnings

Compound Learnings

Transform ephemeral session learnings into permanent, compounding capabilities.

When to Use

• "What should I learn from recent sessions?"
• "Improve my setup based on recent work"
• "Turn learnings into skills/rules"
• "What patterns should become permanent?"
• "Compound my learnings"

Process

Step 1: Gather Learnings

# List learnings (most recent first)
ls -t $CLAUDE_PROJECT_DIR/.claude/cache/learnings/*.md | head -20

# Count total
ls $CLAUDE_PROJECT_DIR/.claude/cache/learnings/*.md | wc -l

Read the most recent 5-10 files (or specify a date range).

Step 2: Extract Patterns (Structured)

For each learnings file, extract entries from these specific sections:

Section Header	What to Extract
## Patterns or Reusable techniques	Direct candidates for rules
**Takeaway:** or **Actionable takeaway:**	Decision heuristics
## What Worked	Success patterns
## What Failed	Anti-patterns (invert to rules)
## Key Decisions	Design principles

Build a frequency table as you go:

| Pattern | Sessions | Category |
|---------|----------|----------|
| "Check artifacts before editing" | abc, def, ghi | debugging |
| "Pass IDs explicitly" | abc, def, ghi, jkl | reliability |

Step 2b: Consolidate Similar Patterns

Before counting, merge patterns that express the same principle:

Example consolidation:

• "Artifact-first debugging"
• "Verify hook output by inspecting files"
• "Filesystem-first debugging" → All express: "Observe outputs before editing code"

Use the most general formulation. Update the frequency table.

Step 3: Detect Meta-Patterns

Critical step: Look at what the learnings cluster around.

If >50% of patterns relate to one topic (e.g., "hooks", "tracing", "async"): → That topic may need a dedicated skill rather than multiple rules → One skill compounds better than five rules