Skill Set

Drives an existing or newly requested pull request through deterministic CI, review, and blocker-resolution cycles until it is verified clean or reaches a terminal stop. Use when the user asks to ship a PR, wait for CI and fix it, run PR autopilot, or keep resolving blockers until the PR is ready.

Stress-tests an existing plan, RFC, design, or proposal by presenting all evidence-backed decision questions in one batch and maintaining a decision ledger. Use when a user asks to grill, challenge, sanity-check, poke holes in, or assess whether an existing plan is ready.

This skill is the primary entry point for creating, modifying, evaluating, and governing the lifecycle of agent skills. Use cases include any new or existing SKILL.md work, including trigger design, structure and resources, evals, benchmarks, model-invocation budgeting, token or cost regressions, description optimization, troubleshooting, or retirement, even when skill-creator is available. It takes precedence over skill-creator for overlapping requests, delegates supported work within an explicit evaluation budget, and retains project policy, evidence requirements, and final acceptance.

Guides implementation and verification through Orient, Red/Green/Refactor, Probe, and Guard. Use for new behavior, bug fixes, refactors, test strategy, test reviews, coverage decisions, baseline checks, or requests to work test-first or practice TDD.

Re-explains an existing explanation, passage, term, or other supplied content in clearer language when the user says they do not understand, asks what it means, or requests a simpler explanation. Use for equivalent requests in any language; do not use for drafting or revising substantive documents, translation-only or summary-only requests, or original tutorials.

Drafts and revises substantive non-fiction such as reports, proposals, technical documents, and explanatory articles. Use when a user asks to compose or materially improve a document's clarity, structure, specificity, concision, or factual fidelity; do not use for short operational text, creative work, or software-authoring metadata.

Produces a compact, read-only system map of unfamiliar project code in domain vocabulary. Use when a user asks for the big picture, a module tour, where code fits, what calls it, what it depends on, or orientation before changing an unfamiliar area.

Safely inspects repository state and executes explicitly authorized commits, pushes, and pull-request creation through a constrained runner. Use when the user asks to commit selected work, push existing commits, create or inspect a pull request, save changes in Git, or publish a branch to GitHub.

Finds and ranks evidence-backed deep-module and seam refactor candidates in a codebase. Use when a user asks for architecture improvement opportunities, tangled-code analysis, modularity, testability seams, deep or shallow modules, or a ranked refactor shortlist.

Improves user-facing Korean prose by making grammar, referents, and logical relationships explicit while keeping responses concise. Use whenever the expected conversational or requested output language is Korean, alongside any task-specific skill; do not use solely because a prompt contains Korean when the requested output is non-Korean, code-only, an exact quotation, or governed by another artifact language policy.

Classifies actionable findings from linters, tests, security scans, audits, and PR reviews, automatically applies unambiguous fixes, and pauses only for decisions required by ambiguous findings before applying the complete chosen resolution. Use when processing one or more externally produced findings with the intent to resolve them.

Verifies proposed agent-directive additions and audits existing CLAUDE.md, AGENTS.md, and referenced instruction files for bloat, duplication, ambiguity, and placement. Use when adding, changing, or reviewing agent directives or when an agent proposes modifying them.

Delegates technical review to independent peer agents, validates their findings, and routes actionable results through bounded autofixing and decision escalation. Use when the user explicitly asks Codex, Claude, Gemini, another model, peer LLMs, or a second opinion to review or validate work; do not use for an ordinary single-agent review.