Agent Claude Code
Attacks the DESIGN of a freshly-authored skill, hook or agent — wrong step graph, missing gate, prose describing a deterministic step, a description that will never fire — and returns triaged findings, dispatched by /author's design-review gate after fill and before verify.
Agent Claude Code
Fresh-context adversarial code reviewer that takes a curated brief (commit SHA range, changed files, project context, optional SPEC) and returns triaged findings — dispatched only when the vendor review chain exits 3 because no independent reviewer answered, and always recorded as NOT independent.
Agent Claude Code
Fresh-context, single-round in-repo investigator, dispatched by /explain and any skill needing a focused deep-dive, that returns structured findings with exact file:line citations so the caller can synthesize without carrying the search traffic.
Agent Claude Code
Flags drift between a SPEC's interpretation and the user's verbatim ask — shape, scope and vocabulary inflation, dropped premises — reading ONLY the user's literal words and the SPEC's behavior contract, dispatched after SPEC writing alongside the adversarial reviewer that hunts holes instead.
Skill Claude Code
Scaffolds a conforming AI-layer artifact — skill, hook, agent or output style — that already passes the workbench's format checks, Anthropic's documented frontmatter shape, and the four token/context/determinism principles. Use when authoring or compressing any skill, hook, agent or output style.
Skill Claude Code
Ends the session — verify what changed, update the project, journal it, land every worktree this thread owns on its repo's trunk, report a line a script proved. Use when the user says "close", "wrap up", "let's close", or any plain-English request to end the session, and when a producer skill's auto-close gate fires.…
Skill Claude Code
Adversarial debate — three agents (Claude, Codex, Grok) argue a question from three positions, then the strongest conclusion is synthesized. Use when the user says "debate this", "I'm torn between X and Y", "red-team this plan", or "council on this".
Skill Claude Code
Deep research into a topic or issue — investigates until confident, then presents an executive summary and root-cause analysis. Use when you need to understand WHY something happens. Takes [topic, question, or issue description].
Skill Claude Code
Adversarial review of freshly-built code — find what was missed, what's inconsistent, what breaks. The single code reviewer for the loop. Auto-fired by /implement (phase-4.7); standalone-callable after any substantial change. Writes an implement-audit: line into the report.
Skill Claude Code
The "do" verb of the loop — execute a SPEC with rigorous self-validation. Use ONLY in a fresh context after running /project in a prior session.
Skill Claude Code needs its repo
Single entry point for project work — renders the live project index, or loads a slug and routes it to the think flow, /implement, /close or a state report, composing /spec, /implement and /close as primitives. Use when the user says "/project", "/project [slug]", "let's work on [project]", "create a project for [X]"…
Skill Claude Code
Adversarial pre-review of any freshly-written or materially-rewritten SPEC or spec folder, running reviewer consensus (correctness) plus the spirit-check agent (fidelity to the user's verbatim ask — spec.md's Input and Clarifications, or a single-file SPEC's § 0 / § 0a — and to the behavior contract) in one pass and…
Skill Claude CodeCodex
See the real running product of any app in your workbench, find its design problems, fix them, and explain what changed. Use after a UI change, before/after committing, when asked to QA, screenshot, fix, or visually verify a web app, or to confirm a portal renders correctly at desktop/tablet/phone. CLI/library targets…
Skill Claude CodeCodex
The vendor review chain the loop skills run and never the model itself — code-review.ts over a git diff, spec-audit.ts over a SPEC or spec folder, policy-review.ts over any artifact — each walking a row of the policy table (policy.json here: Codex first, its declared backup next, never the vendor that wrote the code…
At most 3 mods per repository are shown here, and a mod shipped inside a plugin is left to that plugin's page — the rest are on their repository pages: