skills

Expert system for choosing the optimal Claude model, reasoning effort, and tool permissions. Evaluates task complexity, context bounds, cost constraints, and tool access profiles for Claude Code and native agent skills. Use whenever the user asks to "choose model", "select model", "model effort", "claude model", "claude tools", "claude effort", "configure claude", or asks which model, effort level, or tools to use with Claude.

Systematic root-cause debugging and surgical repair engine. Combines the scientific method (reproduction first, backward data-flow tracing, multi-component boundary logging) with strict anti-cheat and anti-masking verification. Use for "/debug", "debug", "/fix", "fix", "bugfix", "hotfix", "troubleshoot", "root cause", or "investigate bug".

Relentlessly stress-tests and interviews the user on proposed system architecture, database schemas, or feature plans before implementation. Adopts the persona of a Senior Principal Systems Architect. Uses the Design Tree & Frontier Algorithm to batch unblocked questions into rounds with recommended engineering stances. Enforces the Facts vs. Decisions Law. Feeds settled decisions to the configured external SDD skill and records ADRs when repository conventions require them. Use whenever the user asks for "design", "system design", "architecture", "grill me on this design", "stress test my plan", or invokes /design.

Product-specific AI evaluation engine based on Hamel Husain & Parlance Labs methodologies. Diagnoses eval pipelines, discovers failure modes from traces via local review apps, designs code-first evaluators and binary LLM judges, and calibrates alignment using TPR/TNR and Rogan-Gladen statistics. Use for "/evals", "evals", "eval", "ai evals", "error discovery", "trace analysis", "judge calibration", "llm judge", or "evaluate rag".

Review code changes for evidenced correctness, concurrency, DDIA data invariants, failure resilience, craftsmanship, and production risks. Uses an Evidence-Based Judge to reject hallucinations. Use for code reviews, PR reviews, diff inspections, or principal engineering reviews. Trigger with "/review", "review", "code review", or "diff review".

Coordinate specification-driven delivery through a configured external SDD skill, test-first implementation, simplification, engineering review, and evidence-backed handoff. Use for "/ship", "/lifecycle", "full engineering lifecycle", or "build and review this feature".

Authoritative Skill Factory for designing, validating, packaging, and versioning native Claude agent skills. Transforms requirements into standard SKILL.md, deterministic verification scripts, on-demand reference guides, behavioral evals, and distributable packages across host/model matrix. Use whenever the user asks for "/skill", "skill", "create skill", "skill factory", "new skill", "package skill", or "skill matrix".

Rapid, throwaway spike engine designed to answer empirical or "ungrillable" questions raised during architectural design or development. Implements minimal disposable prototypes in isolated scratch workspaces, measures concrete performance or behavior with run_spike.py, and settles technical decisions. Use for "/spike", "spike", "throwaway spike", "timeboxed prototype", "benchmark", "proof of concept", or when an empirical experiment is required.

Enforces strict Test-Driven Development (Red-Green-Refactor) for implementing features, bugfixes, and refactors. Mandates writing failing tests before production code, verifying failure modes, writing minimal production code to pass, and refactoring under green test protection. Use for "/tdd", "tdd", "test driven development", "red-green-refactor", "write tests first", or implementing specs with test-first rigor.