Promotion preflight gate for the agent-skills marketplace. Run before promoting a skill from a sandbox plugin to an official one, or before opening a promotion PR, to get a go / no-go verdict. Audit-first: it chains the deterministic skill-audit signals, marketplace/registry alignment, and version + CHANGELOG presence, and BLOCKS (no-go) on any unresolved finding, citing the failing gate. User-invoked only. For fixing what it surfaces use skill-studio:skill-enhance; for the release wiring use repo-governance:skill-maintainer.
Shapes a loose intent into a well-formed prompt for a Claude Code session that can gather context, use tools, and ask questions. Matches the intent to a validated archetype, fills its slots (objective, context to gather, anchors, clarify-first questions, acceptance criteria, verification, output contract), routes the tier via model-recommender, and self-checks the result with prompt-audit. Use when you have a rough idea of what to ask an agent and want it reshaped to hit a specific goal, when you type /smart-prompt with a loose instruction, or to turn "read X and study Y to figure out Z" into an executable agentic prompt.
Adversarially reviews a draft prompt against a versioned rule set to find what will make it underperform — archetype mismatch, wrong prescription level for the target model, the Fable reasoning-echo trap, negatives that should be positives, unstated expectations, rejected params, and more — then returns tagged findings (rule_id + severity) plus a rewritten, --fix-style prompt. Use when asked to review / audit / critique / improve a prompt, check whether a prompt is good, tighten a system prompt, or before shipping a prompt to a model.
Recommends which Claude model and effort level to run for a task. Classifies the task into an archetype → tier + effort using the rubric in ~/.claude/model-profiles.md, then resolves today's model string from that file's tier_to_model. Use when deciding which model to use, routing a task or phase to a model, picking an effort level, choosing between Opus/Sonnet/Haiku/Fable, or when a task spans plan/implement/review phases that each want a different model.
Audit a Claude skill for quality, context-efficiency, and best-practice compliance — classify its archetype, score the four principles and instruction craft, and return a report — or review a PORTFOLIO of skills for overlap and trigger collisions. Use when someone asks to audit, review, grade, "lint", or sanity-check a skill or plugin's quality; asks "does this skill follow best practices / is it well-scoped / is the description good enough to trigger / do these skills overlap or collide"; or wants a governance sweep across a set of skills. Read-only: it diagnoses and recommends, it does not edit. Do not use to create or scaffold a skill (use skill-create), to improve a skill or run eval/benchmark loops (use skill-enhance), or for versioning/release/CHANGELOG governance (use repo-governance:skill-maintainer).
Post a structured code review to a GitHub PR via `gh` CLI (REST API fallback for multi-comment / inline-anchored reviews). Maps BLOCKER/HIGH/MEDIUM/LOW/QUESTION sections to the right review event and inline anchors, parses ` ```start:end:path ` code refs, and adapts tone to the PR author. Use when you have a written code review and want it posted to a specific PR URL or number. Do NOT use to author the review itself, triage existing comments, or open/merge PRs.
TypeScript code quality patterns -- typed clients with timeouts, Zod input validation, structured domain errors, no any on public APIs, structured JSON logging with pino, OpenTelemetry tracing, and PII redaction. Use when adding external API calls, validating inputs, handling errors, or setting up logging.
Review code, PRs, diffs, or implementation plans for distributed-systems risks, HTTP API contracts, queue consumers, scheduled jobs, and Kubernetes workloads. Focuses on production safety, idempotency, and failure modes.