メニュー
Continuity across long sessions. Snapshots in-progress work before context compaction and restores it after, and nudges auto-compaction toward a natural breakpoint so nothing important is lost mid-task.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
| name | compactor |
| description | Continuity across long sessions. Snapshots in-progress work before context compaction and restores it after, and nudges auto-compaction toward a natural breakpoint so nothing important is lost mid-task. |
Operate as COMPACTOR — continuity across long sessions — for this session.
Before context is compacted, summarized, or the session restarts, snapshot the in-progress work to a durable note:
Keep a durable progress ledger. For any multi-step or multi-task work, maintain .mc/progress.md alongside the snapshot: check it at the start of each step, and append a line the moment each step completes (what shipped + the commit range). Conversation memory does not survive compaction — in real sessions a controller that lost its place re-dispatched an entire sequence of already-finished tasks, the single most expensive failure observed. After any compaction or restart, trust the ledger and git log over your own recollection: if the ledger says a task is done, it's done — don't redo it.
After a compaction/restart, restore from that snapshot and confirm the plan before continuing. Prefer to pause at a natural boundary (a finished step) rather than mid-edit. Tolerate transcript encoding quirks (BOM, odd characters). The goal: never lose the thread of a long task.
Scope & verification discipline for a working session. Triggers when about to claim "done / fixed / passing", when a test is about to be weakened, skipped, commented out or deleted to get a suite green, when a refactor, abstraction or dependency appears that the task never asked for, and on "audit this diff" / "did we actually verify that?". Guidance the model follows — it is not enforcement; nothing here can block a tool call.
The MASTER CLAUDE leader/conductor. Triggers on "master claude", "set up master claude", "mc setup", "onboard me", "build my team", "what's new", or when starting work in a new/unfamiliar project. It interviews you (grill-me), maps the project, then assembles a tailored team from the installed skills and agents — Sentinel the project cartographer plus planning, review, understanding, guardrail and security specialists — and runs it on your work. It builds the best `.claude/` workspace per project, delegates to parallel subagents (choosing the right model per agent), brainstorms hard and decides fast, keeps a complete view of every installed capability, keeps itself and you current with the newest Claude Code features, and proactively offers the right tool the moment a need shows up.
Stay current with Claude Code and Claude itself. Triggers on "what's new", "what can Claude Code do now", "is there a newer/better way", "update master claude", "which model should I use", or before relying on any claim about a Claude Code feature, model, price, or limit. Checks the installed version, the live changelog, and the harness surface MASTER CLAUDE depends on — then reports only the deltas that matter for this project.
How MASTER CLAUDE picks the right model for each agent and task. Triggers on "which model", "pick a model", "model selection", "use haiku/sonnet/opus", "cheaper model", "model per agent", or whenever spawning a subagent/team and choosing its model, or optimizing cost vs. quality on a long run. Gives a role→model heuristic, the "turn count beats token price" rule, the effort dial, and the cost math so the team is both strong where it matters and cheap where it doesn't.
How MASTER CLAUDE delegates work to subagents and teams. Triggers on "delegate", "spawn agents", "parallel agents", "fan out", "run these in parallel", "subagent", "orchestrate", "build a team for this", or whenever a task splits into independent chunks, needs a fresh-eyes review, or means reading many files. Teaches the orchestrator-worker pattern: pick the right primitive (subagent vs agent-team vs background session), write a complete delegation brief, parallelize only when it pays, scale effort to the task, and verify with an adversarial reviewer. Use to plan and run any multi-agent push.
Run parallel or risky work without collisions using git worktrees. Triggers on "worktree", "isolate this change", "try this without touching main", "run these agents in parallel safely", "spike/experiment", or whenever multiple subagents or features would edit the same tree at once. Each task gets its own branch + working directory, so concurrent work never clobbers another's files and a failed experiment is thrown away by deleting a folder. Pairs with subagent-orchestration for true parallel building.
