Manual-only controlled-language modes. Constrains the vocabulary, sentence shape, evidence, and order of what Claude writes to a published rule set, for one target or for the rest of the session. Ships /ste (ASD-STE100), /plain (ISO 24495-1 plain language), /ubiquitous (ubiquitous language), /calibrated (ICD 203 analytic standards), /normative (RFC 2119 keywords), /ears (Easy Approach to Requirements Syntax), /bluf (bottom line up front), /pyramid (Minto), /sbar (clinical handoff), /comments (Conventional Comments), and /diataxis (four documentation modes).
11 skills
Manual-only decision procedures. Each skill runs one published protocol against a judgment already on the table, structuring it without making it. Ships /reversible (one-way and two-way doors), /options (WRAP widening), /tradeoff (Kepner-Tregoe and Pugh), /hypotheses (ACH), /premortem (Klein), /redteam (adversarial review), /fermi (order-of-magnitude estimation), /base-rate (reference class forecasting), /quit (kill criteria), and /adr (Nygard decision records).
10 skills
Verification procedures. Each skill runs one published protocol to produce evidence about code that already exists, rather than an argument about it. Ships /repro (delta debugging), /bisect (predicate-driven bisection), /boundary (equivalence partitioning and boundary value analysis), /differential (differential testing), /property (property-based testing), /mutate (mutation testing), and /characterize (Feathers characterization testing). All are manual except /characterize, /differential, and /mutate, which /hub can run as steps. Three skills bundle a zero-dependency Node script, and /mutate, the only skill that rewrites a tracked file, refuses to run on a dirty tree or a red suite and proves byte-identical restoration.
7 skills · 2 agents
Interrogatives for work in progress: /why /where /what /how /who /when. All are manual except /what, which /hub can run over the changes a run made. Claims carry checkable provenance instead of a confidence score, and nothing here writes.
6 skills
Refinement procedures. Each skill runs one published protocol against work that has been described but not yet defined well enough to execute, driving it down until every unit is unambiguous to a human and dispatchable by an orchestrator. Ships /impact (Adzic impact mapping), /wbs (work breakdown and the 100% rule), /split (Lawrence and Cohn splitting patterns), /examples (Wynne example mapping), /smells (ISO 29148 and requirements smells), and /invest (Wake's INVEST, which emits ready items in the five fields a spoke brief is made of). All are manual except /wbs and /invest, which /hub can run as steps when a task's pieces are not obvious.
5 skills
Manual-only hub-and-spoke orchestration. /hub turns the main session into a dispatcher that decomposes a task, writes self-contained cold-start briefs, gates the plan for approval, fans work out to subagents, verifies what comes back, and integrates. Delegation is the default, and a direct edit is a stated exception the hub has to justify and disclose.
1 skill · 3 agents