madi

Skeptical interviewer for self-grilling a plan/spec in full-autonomous mode (when the user is AFK and won't be answering questions). Pairs with a separately-spawned domain expert sub-agent — the griller asks, the expert defends — and iterates until the spec is sharp. Use when the main agent is in autonomous mode and needs to stress-test a plan without involving the user. Do NOT use for interactive grilling where the user is present — use `/grill-me` or `/grill-with-docs` instead.

Use this agent when you need expert guidance on frontend UI/UX development, including creating new components, improving existing interfaces, implementing design systems, or making design decisions. Proactively use it when creating new UI components or pages, refactoring interfaces for better usability, implementing responsive designs, adding interactive features or animations, ensuring accessibility compliance, or establishing/following design system patterns. Examples: building a new form component with validation and accessibility; cleaning up a cluttered dashboard layout; implementing a complex interaction like a drag-and-drop scheduler.

Plans and executes behavior-preserving refactors (split overweight files, untangle tight coupling, implement spec/ADR-driven structural changes). Use when the caller says "refactor X", "split this file", "implement ADR-NNNN", "clean up the coupling in Y", or hands over an existing refactor plan to execute. Accepts a file path, directory, ADR/spec doc, free-form description, or pre-baked plan. Spawns sub-agents (Explore for callers, verify for tests) as needed. Do NOT use for pure bug fixes, new features, or one-line renames — those are faster in the main agent.

Ad-hoc technical research agent invoked directly by the user (or by griller/expert pairs during /self-grill) to deep-dive a technical topic and return a self-contained, citation-backed briefing. Use for "research X", "what's the best way to do Y", "compare options for Z", or deep questions on auth/OAuth/OIDC/Entra ID/BetterAuth/CASL/B2B-guest-invites/domain-restricted-signup and other libraries/frameworks. Returns a briefing for immediate use AND writes it to the current repo's root .claude/research/ (or the current directory if not in a repo). Do NOT use for GSD pipeline stages (those have dedicated gsd-*-researcher agents), for broad codebase file-location sweeps (use Explore), or for implementation (this agent is read-only).

Use this agent when the user is performing repetitive tasks, manually executing multi-step processes, or when you observe workflow inefficiencies that could be automated or optimized. Also use when the user explicitly asks for workflow tips, productivity improvements, tool recommendations, or ways to work more efficiently. Use it proactively when you notice patterns that suggest optimization opportunities — e.g. the user re-running the same build/deploy sequence by hand, asking for nvim/TypeScript tooling recommendations, about to edit the same config across many files one-by-one, or asking how to structure agents to work more efficiently.

Scaffold a new Claude Code subagent (the kind invoked via the Agent tool). Use when the user asks to "create an agent", "make a subagent", "add an expert for X", or "/agent-creator" — OR when, mid-task, you notice a recurring specialised job (e.g. dense log triage, security review, migration planning) where having a dedicated agent would help on this and future runs. Writes the agent file to `~/.claude/agents/` (global) or `<repo>/.claude/agents/` (project).

Establishes the operating discipline for a fully-autonomous run where the user is AFK and will NOT answer questions. Invoke when the user says "go autonomous", "work autonomously", "full autonomous mode", "don't stop to ask me questions", "/loop", or otherwise signals they're stepping away and you must drive to completion without them.

Ultra-compressed communication mode. Cuts token usage ~75% by dropping filler, articles, and pleasantries while keeping full technical accuracy. Use when user says "caveman mode", "talk like caveman", "use caveman", "less tokens", "be brief", or invokes /caveman.

Identifies and deletes session-generated artefacts (planning docs, specs, self-grill outputs, code review working docs, .claude/specs) that were created during development but have no value in the merged codebase. Invoke when the user says "clean up docs", "remove slop docs", "clean up before merge", "tidy up the branch artefacts", or "clean up the session docs".

Disciplined diagnosis loop for hard bugs and performance regressions. Reproduce → minimise → hypothesise → instrument → fix → regression-test. Use when user says "diagnose this" / "debug this", reports a bug, says something is broken/throwing/failing, or describes a performance regression. Pass "tdd" (/diagnose tdd) to make the failing regression test mandatory and fix under the tdd skill.