core

Non-author verification expert — receives a thin mandate, fetches its own evidence via Linear MCP tools directly, judges independently, posts its verdict directly or returns it to the caller. Eleven mandate types under playbooks/. Invoked as a fresh spawn whenever work needs independent verification before it reaches Done or the operator. Triggers on a caller spawning `@attack-kitty` with a mandate, or programmatic invocation.

Fires before the first Agent, Task, or Workflow call of a piece of work — about to spawn a subagent, fan out, parallelize, split work across contexts, or run one problem as several attempts. Skipped, delegates ship dead-on-arrival work from vague briefs. Consulted, it picks the execution shape (including not delegating at all) and gives each brief a checkable done-condition.

Build and sharpen a project's domain model — pin down terminology, challenge fuzzy language, record architectural decisions. Works for engineering, strategy, research, and design domains.

Interview-style stress-test for a plan, decision, or idea — one question at a time, environment facts looked up rather than asked, decisions put to the operator with a recommended answer, no action until shared understanding is confirmed. Triggers on "grill me", "stress-test this", "let's align", "grill this plan/decision/idea", or whenever the operator wants her thinking pressure-tested before acting.

CLAUDE.md orientation expert — read and write the Re-entry Cue + frontmatter routing fields (linear_project_id, knowledge_intake, description, status). Invoked by /session-start, /session-closeout, and /new-project. Triggers on "/project-state read", "/project-state write", or programmatic invocation.

Build a throwaway prototype to answer a design question. Use when the user wants to sanity-check whether a state model or logic feels right, or explore what a UI should look like.

Research discipline for exploratory search, purchase research, and primary-source investigation. Carries search methodology (mode detection, funnel stages, validation, retrieval tactics) and routes to playbooks by search type. Triggers on research tasks, shopping, purchase decisions, "find me", "what should I buy", "research this", exploratory searches, evaluating products/options to inform a decision, when the user shares a link/article/video asking "what do you think of this" or "analyze this," or invoked as "/research ticket <issue-id>" to resolve a research ticket.

Triggers when the user says "Close out this session for [project]", "Session closeout", "Close this session", or similar closeout phrases. Also triggers on "/session-closeout".

Triggers when the user says "I'm working on [project]", "Let's work on [project]", "Starting a session on [project]", or similar session-start phrases. Also triggers on "session start" or "/session-start".

Correctness layer for lifecycle transitions — verifies tickets are well-formed, receipts are legitimate, and transitions are earned, then executes them directly via the fused script path. Invoked at the point a ticket or map needs to move through claim, begin, mark_done, park, block, un-park, cancel, or close-map.

Cut a body of work too big for one session into vertical slices — complete, usable increments that reach through every layer, not horizontal stages. A first-principles method, five principles each taught by the failure it prevents, fired at the first cut and every re-cut. Domain-agnostic — software, a launch, a research program, an org change. Invoked when decomposing work into slices — `/wayfinder` calls it at charting and building; any planning context can call it directly.

Chart and work a map — turn a loose idea too big for one session into decision tickets on Linear, resolve them one at a time with the operator, then build from the operator-confirmed intent through validated slices. Domain-agnostic — software, strategy, content, research. Invoked by the operator — "chart a map", "work the map", or a named map.