メニュー
Run the verifier alone — a separate clean-context agent that returns PASS/FAIL against goal.md (not the agent that wrote the code). Use when Malik says 'verify it', 'is it actually done', 'check against the goal', or 'did this pass'. Tests passing ≠ goal met.
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
Health-check the memory bank, rules, and index — flag stale, contradictory, or redundant entries and spec↔code drift. Use when Malik says 'audit the bank', 'is the memory stale', 'check spec harness health', or 'clean up the rules'.
Run the SDD pipeline over tasks.md: implement → test → VERIFY → review, one task at a time, under the ratchet. Use when Malik says 'build it', 'run the pipeline', 'implement the tasks', or 'ship this feature'. Each task gated by the separate verifier; lessons harvested by learn.
Bind the 6 SDD agents to THIS project's rules — fill each agent's GEN:rules region with its concern's real rules (implementer←Coding, tester←Testing+commands, reviewer←Reviewing, researcher/planner←Architecture, verifier←goal). Use when Malik says 'generate the agents', 'make the agents project-specific', 'bind agents to the rules', or right after rules are created/updated. If no rules exist yet, it routes to rules-generation (3 sources: structure + project best-practice + the internet) first. Idempotent; re-run after learn updates a rule.
Add or refresh Spec Harness in a repo — greenfield (new) or existing (integrate), idempotent. Use when Malik says 'install spec harness', 'set this repo up with spec harness', 'integrate spec harness into X', or 'start a new spec-harness project'. Runs the interview (mode, stack, constraints, work-intake source, learning goals, regression guards) and scaffolds the bank + agents + commands + skills without clobbering existing CLAUDE.md/AGENTS.md.
The learning loop — turn feedback, a correction, or a verifier-FAIL root cause into one enforced rule injected into the ratchet (AGENTS.md/ai_rules), dated. Use when Malik says 'learn this', 'remember not to do X', 'capture this lesson', 'don't do that again', or right after he corrects you. Closes the loop the verifier opens; emits an optional human lesson too.
Turn a spec into a design.md (the 'how' — architecture decisions). Use when Malik says 'plan the implementation', 'how should we build this', or 'design this'. Second SDD step.
| name | spec-harness-verify |
| description | Run the verifier alone — a separate clean-context agent that returns PASS/FAIL against goal.md (not the agent that wrote the code). Use when Malik says 'verify it', 'is it actually done', 'check against the goal', or 'did this pass'. Tests passing ≠ goal met. |
Spawn the sdd-verifier with a clean context. It runs each goal.md end-state check literally (no inference) and emits RESULT: PASS|FAIL + evidence. It must not be the agent that wrote the code.
.claude/commands/spec-harness/verify.md (this repo) or the spec-harness source commands/verify.md..claude/agents/: verifier..memory/) · Spec-Driven Development · Harness (ratchet AGENTS.md + verifier + learning loop). Read SPEC-HARNESS.md for how this repo is wired.goal.md.spec-harness-learn → a dated rule.