claude-workspace

Maintains a clean, predictable project file structure. Moves, renames, and creates directories via shell — never edits the content of any file. Returns the updated structure map so the orchestrator can record .workspace/file-structure.md.

Rebuilds project state when context is lost or drifting. Re-reads the canonical artefacts (proposal, tasks, file structure) and the actual tree, then reconstructs a faithful mental model. Read-only — it changes nothing and reports only what IS.

Adversarially checks an implementation against its task spec. Read + run-tests only (no Edit/Write) — it detects deviations and never fixes them. Outputs pass/fail, deviations with evidence, and required corrections for the implementer.

Executes a single task exactly as written in its task spec. Fresh session, no planning, no reinterpretation, no creativity — only execution. Full read/edit/write/bash access. Stops and reports rather than guessing on ambiguity.

Turns a raw idea (plus any research brief) into the canonical proposal — problem, constraints, invariants, acceptance criteria, scope boundaries, required artefacts. The root of truth for the whole system. Writes only the active initiative's proposal.md.

Work through a hard or consequential choice and reach a clear, defensible recommendation. Use when the user is weighing options, stuck between alternatives, facing a tradeoff, or about to commit to something costly or hard to reverse. Frames the real decision and widens the option set, evaluates options against weighted criteria, adversarially checks for motivated reasoning and false binaries, then recommends — with the tradeoff being accepted made explicit. Built to defeat the pre-chosen answer dressed up as analysis.

Generate a wide, varied, genuinely original set of ideas for a problem, opportunity, feature, name, angle, or open question — then sharpen the best ones. Use for brainstorming, exploring alternatives, "how might we", finding lateral or non-obvious approaches, or when the user is stuck in one frame and needs range. Frames what makes an idea valuable, diverges hard across multiple lenses, adversarially checks the set for anchoring and missing categories, then converges on the most promising. Built to defeat premature convergence.

Turn a vague idea or rough problem statement into a crisp, testable concept and a per-hypothesis test plan, captured as a durable ideation.md artefact. Use before research or proposal-writing, when the concept itself is not yet decided — "ideate on this", "spiral this", "let's refine the idea", "help me figure out what to build", "I have a rough idea, not sure what it should be yet", "explore this before we spec it". Runs an iterative Frame → Diverge → Challenge → Converge loop with an objective structural-completeness gate that blocks premature convergence. Self-contained: no dependency on other loose local skills.

Turn a goal, problem, or vague intent into a stress-tested, actionable plan. Use when the user wants to plan an initiative, project, feature, decision, migration, or any effort — especially when the objective is fuzzy, the path is uncertain, or constraints compete. Runs a Socratic intake to pin down the real objective and constraints, drafts a structured plan, then adversarially pre-mortems it before finalizing. This is a thinking process, not a template-filler.

Look back on completed or in-progress work, a decision, or a stretch of effort and extract what actually happened and what to genuinely learn. Use for retrospectives, post-mortems, "what went wrong / what went right", reviewing a finished project, or when the user wants to think clearly about a past experience instead of rationalizing it. Reconstructs events without spin, builds a structured retrospective, then adversarially challenges the comfortable narrative before distilling concrete changes.

7 hooks across 3 events