precompact
Precompact
Persist what you know to durable storage. Memory files should reflect reality — not a stale snapshot from three tasks ago.
On Activation
Work through these steps. Skip any that don't apply.
1. Update memory files
Review your auto-memory directory and update it:
- Active work: branches, commit hashes, what's done vs pending, current branch relationships
- Decisions made: anything decided in this session that isn't written down yet
- Corrections received: if the user corrected you, memory must reflect the correction
- Remove stale entries: merged PRs listed as open, old branch names, completed tasks, outdated notes — fix them
2. Clean up task list
If tasks were used in this session:
- Mark completed tasks as completed
- Remove tasks that are no longer relevant
- Update in-progress descriptions if scope changed
3. Persist undocumented decisions
Scan the conversation for things decided but not yet written anywhere durable:
- Naming conventions agreed on
- Architectural choices made
- Patterns established
- User preferences expressed
Write these to the appropriate memory file.
4. Report what you did
Arguments
If the user passes arguments, treat them as explicit instructions to persist before doing the rest.
Do
- Be thorough — if you know it and it matters, write it down
- Remove stale information (wrong is worse than missing)
- Update hashes, branch names, PR statuses to current values
Don't
- Create new files unless memory genuinely needs a new topic
- Ask questions — persist what you know, not what you don't
- Touch code or make changes beyond memory/task housekeeping
More from corygabrielsen/skills
mission-control
Coordinate complex multi-step work using task graphs and parallel background agents. Use when work requires decomposition, delegation, and long-running operations that may survive context compaction.
11loop-codex-review
Automated code review loop with progressive reasoning levels. Runs n parallel Codex reviews (configurable via -n), Claude addresses issues, climbs from low→xhigh reasoning until fixed point (all n clean). Human approval at each iteration.
11review-pr
Thorough, educational PR review process. Gathers context, explains changes, identifies issues systematically, and drafts constructive feedback collaboratively before submitting.
11orthogonalize-pr
Identify and separate orthogonal (independent) change sets within a decomposed PR. Creates branches for each set that could theoretically be reviewed or merged separately.
10decontextualize
Course-correct when artifacts overfit to conversation context. Invoke when examples, names, or language are too specific to current work.
10loop-review-skill-until-fixed-point
Iterate /review-skill on a target until fixed point. Runs review passes until all reviewers return NO ISSUES.
9