sigma

Check or scaffold CLAUDE.md / CLAUDE.local.md against best-practice research (official Anthropic guidance + community consensus + real-world examples). Use when running /claude-md-check or /claude-md-create, when a human asks to "check", "audit", "improve", or "review" a CLAUDE.md file, or when scaffolding a new one and native /init would produce unbounded, undisciplined output. Carries the length/structure/specificity rubric both commands share.

Estimate, measure, and optimize token cost for sigma's heavy operations (review's three axes, the profile walk, loop cycles, multi-model research). Use before running a heavy op to size it and pick model tiers, and after to record actual spend and report trends. Triggers: "how expensive", "estimate cost", "which model should this axis use", "what's burning tokens", "sigma cost", or before any multi-axis review / loop / deep research run.

Generate EXTERNAL-facing project documentation from the codebase itself — README, API/reference docs (from code signatures + docstrings), CHANGELOG / release notes (from git log), and an "about this project" presentation. Grounds in ARCHITECTURE.md when present; delegates HTML/theming for presentations to sigma-present. Use when the user wants docs for an AUDIENCE (users, contributors, release readers) — not internal onboarding. Triggers: "write a README", "generate API docs", "draft release notes", "make a project overview deck", "document this for users".

Use when refining already-verified code to fight AI slop — over-abstraction, dead code, duplication, needless cleverness — while preserving behaviour exactly. Powers `sigma loop --simplify`. The simplifier is a DISTINCT agent from the implementer; cleanup is re-verified, never a gate.

Find loaded-but-unused MCP servers and plugins that bloat every Claude context, and reversibly disable the dead weight to save tokens + cost. Use when the context feels heavy, tool lists are huge, or you want to trim what's loaded. Triggers: "too many tools", "context is bloated", "what MCP/plugins am I not using", "save context", "prune unused", "sigma prune", or a periodic context-hygiene pass.

Discover skills relevant to sigma's domains on skillsmp.com and keep the bundle fresh, without ever auto-installing. Use to find trending or new community skills for classic-ml / deep-learning / nlp / rl / data / agents / mlops / llm work, vet them, and pull the keepers in. Triggers: "find new skills", "what skills should sigma add", "scout skillsmp", "keep the bundle current", "any trending skills for <domain>", or a periodic bundle refresh.

Adversarially grill a sigma design (architecture.md) or specification (spec.md) BEFORE the pipeline advances to code — catch logic flaws, ambiguity, untestable acceptance criteria, missing edge/error paths, scope creep, and ML/data risks while they still cost a sentence to fix instead of a rewrite. Use when running /grill, when a human or the loop wants to pre-flight a spec or blueprint, when about to decompose a spec into tasks, or when the user asks to "review", "pressure-test", "stress-test", "red-team", or "poke holes in" a design/spec. Carries the interrogation rubric; the griller is a separate agent from the author.

Run a bounded auto-grill loop over a sigma spec or blueprint — grill, triage findings into mechanical-auto-fix vs human-surface, let a DISTINCT editor agent apply only the mechanical fixes, re-grill, and repeat until READY or a round cap / no-progress stop. Use when running /grill-loop, when a /grill pass came back BLOCK with many fixable findings and you want them driven down without hand-editing every one, or when the user asks to "auto-fix the spec", "iterate the spec to READY", "grill and edit in a loop", or "converge this spec". Keeps editor ≠ griller (no self-grading) and never auto-edits intent.

Recall sigma's past ratcheted lessons before implementing or verifying a task, so previous mistakes are not repeated. Use when about to write or review code in any sigma domain (classic-ml, deep-learning, nlp, rl, data-analysis, data-engineering, ai-agent-engineering, mlops, llm-engineering), when starting a loop task, or when the user asks "what have we learned" / "avoid past mistakes". Lessons are written by the loop on failure and by /sigma-learn-lesson; this skill reads them back.

Auto-surface the right sigma domain context-engine when implementing or verifying a data-science / AI-engineering task. Routes to hand-authored implementer guidance, verifier checks, and a logic-evaluator for the matching domain. Use when a task touches: classical ML (scikit-learn, feature engineering, cross-validation, model selection), deep learning (PyTorch / TensorFlow, training loops, architectures, distributed training), NLP (tokenization, transformers, classification, NER, semantic search, summarization), reinforcement learning (policy gradient, actor-critic, value-based, reward shaping, RLHF, environments), data analysis (EDA, A/B testing, statistical testing), data engineering (Airflow DAGs, dbt models, Spark jobs, pipelines), AI agent engineering (agent harness design, tool definitions, orchestration, evals), MLOps (experiment tracking, model registry, monitoring, production readiness), or LLM engineering (prompt engineering, RAG, fine-tuning, LLM evals). Trigger when about to write or review code i

Export sigma markdown artifacts (research.md, spec.md, plans, verify results, kanban board.md) as a polished, single self-contained HTML file — either an interactive reveal.js slide DECK or a long-scroll animated REPORT page, with an optional kanban snapshot. Use when the user wants to share, present, email, or hand off a sigma artifact as a standalone .html (or print it to PDF) rather than as raw markdown. Triggers: "turn this spec into slides", "export research as a report", "make a deck from the board", "share these verify results".

You MUST use this before any creative work - creating features, building components, adding functionality, or modifying behavior. Explores user intent, requirements and design before implementation.

Use when encountering any bug, test failure, or unexpected behavior, before proposing fixes