Step-by-step article writing workflow with style selection (personal voice vs serious), a hallucination-prevention fact-check gate, and AI-smell proofreading. Use when writing or co-writing blog/tech articles — triggers: 「記事を書く」「ブログ記事」「記事化して」, turning memos/logs/repos into an article, or drafting for funailog/Zenn/note. For proofreading-only requests, use the japanese-ai-writing-proofreader skill directly.
Proofread Japanese prose in three passes: textlint (mechanical), AI-smell removal, and deep naturalness (sentence-end variety, rhythm, connectives). Use when asked to 「校正して」「推敲して」「AI臭を消して」「自然な日本語にして」, when polishing article drafts / PR bodies / docs written in Japanese, or as the proofreading phase of the article-writing skill.
Iteratively improve agent-facing text instructions (skills, slash commands, task prompts, CLAUDE.md sections, code-generation prompts) by having a bias-free executor run them and evaluating from both sides (executor self-report + caller-side metrics). Repeat until improvement plateaus. Use immediately after creating or significantly revising a skill/prompt, or when unexpected agent behavior is suspected to stem from ambiguous instructions.
Design where a repository's agent knowledge lives — which conventions belong in which directory-level CLAUDE.md, and which "convention-definition" skills should be copied into CLAUDE.md and then deleted. Optimizes for what the model gets wrong without the note, not documentation coverage. Loads the session model's profile from ~/.claude/references/model-profiles/ to match its writing style. Use when onboarding a repo to Claude Code, after a model upgrade, or when CLAUDE.md has grown bloated or conventions are scattered across skills. Triggers: 「CLAUDE.md を設計」「階層別 CLAUDE.md」「作法を CLAUDE.md に移したい」.
Use when the user wants to set up a recurring or autonomous agent loop and needs to design it before running it. Triggers: 「ループを設計」「自律ループを 作りたい」「loop を組みたい」, turning a vague "I want a loop that does X" into a concrete spec, or choosing a runner (/loop, /schedule, ralph-loop, pr-monitor) for repeated work. NOT for running an already-designed loop, and NOT for one-off tasks.
Ship a GitHub Issue end-to-end: fetch the issue as the spec, implement via a worktree-isolated subagent, open a PR that closes the issue, then watch CI through merge. GitHub Issues is the single source of truth.
Watch a GitHub PR asynchronously until CI is green and review threads are resolved (or through merge), using the harness Monitor tool so the main session keeps working. Optional auto-fix mode reacts to CI failures by fixing and pushing on the PR branch. Use when (1) the user asks to watch a PR or CI (「PR見てて」「CI監視して」), (2) changes were just pushed and CI needs confirmation, (3) another skill (e.g. issue-ship) needs CI watched through merge. Self-contained — no plugin dependency, so it works on every machine.
E2E testing workflow for web applications using Chrome in Claude. Automates functional testing, bug detection, and GitHub Issue creation. Use when: (1) running functional tests on web apps, (2) verifying staging/production behavior, (3) finding bugs and creating Issues, (4) capturing UI behavior with screenshots