Extract and verify implicit assumptions before implementing. Use before any non-trivial implementation, bug fix, integration, or refactor - especially when the task relies on undocumented behavior, says "should" or "probably", or touches code you haven't read yet.
Structured diff review with persistent reviewer preferences. Use when asked to review code, review a PR or diff, check changes before commit, or whenever the user states a review nit, style rule, or correction worth remembering.
Preserve the minimum sufficient working state required for another coding agent to continue a task without repeating repository discovery, reconstructing prior decisions, or losing task-specific constraints.
Frontend and UI work discipline. Detects and reuses an existing design system before writing any interface code, and applies sound UI craft when no system exists. Use whenever creating or modifying components, pages, layouts, themes, styling, or any visual interface.
Batched git commit workflow. Use whenever asked to commit, write a commit message, or save work. One shell call gathers repository state, one silent reasoning step composes the message, one shell call commits; Claude only where needed.
Write like a person. Use when drafting or editing anything meant to read as human writing such as READMEs, documentation, blog posts, commit messages, emails, announcements, or changelogs. Also use when asked to make text sound less AI-generated, more natural, or more human.
Test-driven development loop with enforced verification. Use when implementing new behavior, fixing bugs with reproducible symptoms, or whenever the user asks for tests, TDD, or test coverage.
Learns how a specific GitHub reviewer reviews, from everything they've written on a repo's PRs (inline comments with the code they point at, review summaries, conversation), and saves their rules with real examples so code-review can review like them. Use when asked to learn a reviewer's style, review "like ", audit…
Makes risky changes to existing code without breaking callers or data. Pins current behavior with characterization tests before a refactor, and splits schema, API, config, and dependency changes into expand, migrate, contract steps that can each ship alone. Use for refactors, renames across modules, database…
Teaches the user the key invariants of a codebase, module, PR, or concept, so they can reason about it themselves, using prediction questions and recall checks instead of a lecture. Use when the user says "teach me", "help me understand", "what do I need to know about", "explain how X works so I get it", or wants to…
Put a CLAUDE.md on a diet and structure repo docs so an agent reads what the task needs and nothing else. Use when a CLAUDE.md is over 8k tokens, when doctor flags one, when asked to write or reorganize CLAUDE.md or agent docs, or when session notes keep landing in CLAUDE.md.
Structured research loop with pre-declared stopping criteria, a disconfirming probe, and a per-repo hypothesis map. Use for any research task, literature or web search, "find out whether / is it true that" questions, strategy or hypothesis hunting, and autonomous research loops (keep looking, find something that…
Interviews the user and writes a short spec before building a feature that touches several files or has unclear requirements, so the build matches what they actually need and how it will run in production. Use when asked to build, add, or implement something non-trivial, when the request leaves out users, inputs…
Proves a branch or PR works the way it will run in production, feature by feature, and fixes what fails until every claim has evidence. Use before saying a feature is done, before opening or merging a PR, when asked to "QA this", "test everything", "make sure it works", "will this work in prod", or after a change that…
★not rated 2 4d agoA95 tokens
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: