Agent Claude Code
Adversarial QA in two modes. Mode A (default) on a change that already passed its existing tests. Writes new tests aimed at breaking it, runs the code, adds temporary logging to observe real behavior, checks the diff against the ENTIRE pasted ticket or the full set of the user's actual quotes, and hunts for provable…
Agent Claude Code
Repairs factual damage in a paraphrased, translated or heavily rewritten text, using the smallest possible edit and always in fresh wording. Use after a transformation that improved how the prose reads but may have altered what it says, such as a back translation round trip or an aggressive rewrite. Takes the original…
Agent Claude Code
Only runs when bad-cop found something Critical or High. Takes bad-cop's findings, researches the root cause and the correct fix, applies it, and gets every test green (bad-cop's new adversarial tests plus the existing suite). Reads the ENTIRE ticket or the full set of the user's quotes, never a summary, and confirms…
Agent Claude Code
Cheap claim checker. Given a claim that something is done, working, finished, or covered, it reads the actual code and reports whether the claim is true. Non blocking, stamps nothing, never satisfies any gate. Dispatch it liberally and backgrounded whenever you say you finished something, including a plan or a spec…
Agent Claude Code
Understands the codebase and the general engineering standard for a change before anyone touches it. Runs clean-rag's own index, vector search, and import graph (triggering /index-project itself when a project is stale or missing), plus the manual GraphRAG layer for deeper cross file questions, and researches what…
Agent Claude Code
Swipes working code instead of writing it from scratch. Hates original implementation. Checks whether the thing already exists in the project, the stdlib, an installed dependency, the skills.sh skill registry (for "make the agent do X" tasks), or on GitHub/StackOverflow, and if it does, hands back the exact command or…
Skill Claude CodeCodex
Spawn bad-cop directly on a diff that already passed its tests. Adversarial QA, writes tests aimed at breaking it, runs the code, adds logging, reports real provable failures. Reports only, fixes nothing. Stamps VERIFIED itself when it finds nothing, emits NITS when everything it found is nit severity (you fix those…
Skill Claude CodeCodex
How researcher builds a structural picture of a codebase and checks a change against real engineering standards, before swiper ever looks at what to swipe. Covers the quality lenses (correctness, security, testing, maintainability), curating the import graph instead of dumping it, and when to reach for the deeper…
Skill Claude CodeCodex
How to pick a debugging technique instead of guessing, and the tools and CLI recipes for every stack. Read this when a bug is not yielding after two or three attempts, when you are about to add another print statement, or before debugging a stack you have not touched before (React Native, .NET, native, browser). Also…
Skill Claude CodeCodex
Gather official documents (statutes, regulations, any authoritative source) on any topic, chunk them at real citation boundaries, and build a persistent, searchable corpus inside clean-rag. Topic agnostic -- the schema is generic, your topic is whatever source list you write. Use to ingest a new topic's sources or to…
Skill Claude CodeCodex
WHEN doing browser QA, fixing UI issues, or when users express dissatisfaction with visual appearance or behavior; use Playwright MCP to capture screenshots and run a structured visual feedback loop. USE before calling any UI change done.
Skill Claude CodeCodex
Spawn good-cop directly, only after /bad-cop found something Critical or High. Researches the root cause of bad-cop's findings, fixes it, gets every test green, and stamps the verifier gate with VERIFIED. Skip this entirely when bad-cop emitted VERIFIED (it found nothing) or NITS (nit severity only, which you fix…
Skill Claude CodeCodex
Build or query the semantic code graph (true GraphRAG) for a project. An LLM reads the codebase and extracts entities plus the intent behind calls, imports, and dependencies, then answers cross file "how does X flow to Y" questions the import graph can't. Manual and resumable: a build runs overnight and can be killed…
Skill Claude Code
A relentless interview to sharpen a plan or design. Ask the whole answerable frontier at once, each question with a recommended answer, and send researcher and swiper after any fact instead of asking the human for it. Ends with a mermaid diagram of what got settled and what is genuinely still open.
Skill Claude Code
Grill the user relentlessly about a plan, decision, or idea. Use when the user wants to stress-test their thinking, or uses any 'grill' trigger phrases.
Skill Claude CodeCodex
Audit and rewrite prose so it reads as human written and is easy to understand. Use when asked to remove AI-isms, clean up AI writing, make something sound less like AI, or make an explanation clearer. Three modes: detect only, rewrite, and edit a file in place. Covers two separate problems, register (does it read as…
Skill Claude Code
Run the /workshop adversarial loop on a PLAN or a decision, with counterparty safety rules on top. You write the plan, bad-cop attacks it, good-cop revises it with cited grounding, and the loop repeats until bad-cop stamps VERIFIED on a clean pass. Use for any consequential non code decision, an email or letter going…
Skill Claude Code
WHEN the user asks for a slide deck, a presentation, a PowerPoint, a .pptx, or asks to turn an explanation of any topic into slides; build the deck into the active workspace with python-pptx, render every slide and actually look at it before handing it over, then open it. Also covers narrating a finished deck to an…
Skill Claude CodeCodex
Quick mode for this turn. Use when you (the human) already know a change or question is trivial and needs no research or verification. Starting a turn with /ps deterministically stands down the research gate and the verifier for that turn, so a rename, a one line fix, a comment tweak, or a quick question skips the…
Skill Claude Code
Spawn quick-cop on a claim. Cheap check that you actually did what you just said you did. Non blocking, stamps nothing, never satisfies the verifier gate and never replaces bad-cop. Use it liberally, backgrounded, every time you say something is done, finished, complete, or has no gaps.
Skill Claude CodeCodex
How a research agent decides what to search, where to search it, and how to turn a project's import graph into something useful. Covers depth vs breadth routing, the does-this-already-exist check, and per aspect coverage.
Skill Claude CodeCodex
Send out the research agent on a question, with no code edit involved. Use when you want depth and breadth research on a decision, a design question, a "what's the best way to X", or "does this already exist" before committing to anything.
Skill Claude CodeCodex
Spawn researcher directly on a task, no /start sequence attached. Use for a pure codebase structure or standards question, what breaks if this changes, what does good code for this look like here, without needing swiper's swipe check afterward.
Skill Claude Code
Mine the project's memory store for recurring issues, then SWIPE an existing published skill that solves each one instead of writing a new skill by hand. Use when the user asks to turn memory, feedback, or recurring issues into skills, to find a skill for something that keeps coming up, or to check whether a published…
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: