メニュー
Router over the harness skills — which skill fits your situation, and the main flow through them. NOT for performing the routed work itself — it only points; invoke the skill it names.
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
Dispatch a multi-wave plan to specialist agents with audit + risk-area abort. Supports --auto-push, --auto-merge, --goal-mode for budgeted runs. NOT for writing the plan itself (that is /spec), and NOT for a single small edit with no waves — just make the edit.
Independently re-verify a completion claim in a separate context — mechanical evidence (files/tests/assertions) plus a refute-by-default semantic judge — before a wave or task is allowed to be called "done". Emits the shared verdict schema. NOT for code-style review (that is a reviewer's lane) and NOT for verifying work you authored in the same context — the verifier must be a fresh spawn.
Upstream planning-discipline — brainstorm a feature into spec.md + plan.md under .agent/plans/<slug>/, then route to plan approval. Enforced by the spec-gate tool boundary, not prompt coercion. NOT for trivial/test/doc-only edits (the gate scopes those out), and NOT for executing an already-approved plan (that is /supervise).
Commits staged changes and opens a PR. Runs gitleaks + risk-area guards before committing. Push is user-confirmed by default; --auto-push and --auto-merge are opt-in. NOT for review or verification (run those before wrapping), and NOT when nothing has changed yet — there is nothing to commit.
Agent-driven, read-only self-audit of the harness — run the machine integrity layer once as a dry-run, present a per-check pass/fail table, cite the P1-1 doc-reality result, and for any failure give root-cause + fix + backlog follow-up. Consumes the machine gates; it does not reimplement them. NOT for auditing a consumer project's own codebase (use that project's test suite), and NOT for applying fixes — this skill observes and reports only.
Survey a codebase area before changing it. Maps file tree, traces dependencies, identifies idioms.
| name | harness-help |
| description | Router over the harness skills — which skill fits your situation, and the main flow through them. NOT for performing the routed work itself — it only points; invoke the skill it names. |
| disable-model-invocation | true |
| tools | Read |
You don't remember every skill, so ask. Most work travels one main flow;
everything else is standalone. Each skill below is documented in its own
skills/<name>/SKILL.md — this router only says when to reach for which.
/spec <slug> — turn a fuzzy request into spec.md + plan.md under
.agent/plans/<slug>/, then route to plan approval (ExitPlanMode). Add
--interview when the request is fuzzy enough that a wrong guess would
commit the spec to the wrong shape. Approval is what unlocks editing —
spec-gate stops asking once the plan-approval flag exists./supervise <slug> — run the approved plan wave by wave, dispatching
specialists per the delegation contract and auditing after each wave.
--goal-mode for budgeted, resumable runs./verify-completion <claim> — before accepting any "done": a fresh
context re-checks the claim mechanically, then a refute-by-default judge.
/supervise calls this at its completion step; reach for it directly
whenever a builder's word is the only evidence./wrap — commit + PR with the safety gates (gitleaks, risk areas)
intact. --auto-push / --auto-merge are opt-in escalations.A single small edit needs none of this — just make the edit; the gates scope trivial work out by design.
/harness-audit — read-only health check of the harness itself: one
verify-all.sh dry-run, a per-check table, the doc-reality verdict named.
Run it before a release or after a structural change./project-init (command) — scaffold a consumer project with
CLAUDE.md, rules, and gitleaks.toml.Skills are the methodology; hooks are the enforcement. When one interrupts you mid-flow, it names its own escape:
/spec.Any skill added to or removed from skills/ updates this file in the same
commit — a router that lies is worse than no router.