compound-simple

compound-simple

Simplified compound skill for Fast Model and similar models. Embed session learnings in <3 turns.

When to use

At the end of every autonomous work session, after completing work but before commit.

Procedure

Execute these steps in order:

Step 1: Review what changed

Run git diff --stat HEAD~N..HEAD where N is commits this session. Answer:

• What files changed?
• What was accomplished?

Step 2: Check for learnings (4 questions)

Answer each question. If yes, note what to update:

1. Non-obvious fact discovered? (API quirk, hidden constraint, unclear convention) → Update relevant file (CLAUDE.md, skill, or project file)

2. Failure mode others should avoid? (silent error, misleading config, common mistake) → Add warning/gotcha to relevant skill

3. Technique that worked well? (debugging strategy, analysis pattern) → If generalizes: note for future gravity evaluation

4. Convention friction? (rule didn't fit, schema too rigid) → Update convention or note friction

Step 3: Check for implied tasks

If you completed an experiment/analysis, scan EXPERIMENT.md Findings for:

• "N too small" → task: larger replication
• "FAIL" or "below threshold" → task: refined experiment
• "unexpected", "mechanism unclear" → task: diagnosis
• Multi-phase plan → check phase tasks exist

Step 4: Act on findings

For each compound opportunity:

Type	Action
Small update (1-3 lines)	Apply directly to file
Larger change	Create task in TASKS.md
Recurring pattern	Create task: "Run /gravity on: "
Governance change	Write to APPROVAL_QUEUE.md

Step 5: Report actions

Output in this format:

Compound: N actions
- <action 1>
- <action 2>
...

If no actions: Compound: no actions this session.

What NOT to do

• Do NOT run full /compound skill — too complex
• Do NOT start new work — compound embeds learnings only
• Do NOT scan cross-session patterns or convention registry
• Do NOT make governance changes without approval queue
• Do NOT add >10 lines without identifying lines to compress