dotfiles

Generic thinking protocol for deriving optimal solutions through iterative self-checking. Invoke BEFORE searching within the given problem space whenever the user asks for advice, judgment, evaluation, or choice. Reframes the problem, generates options, evaluates trade-offs, runs a self-answer check (does this answer the user's actual question at the right granularity, reach upstream, match expectations, offer non-obvious value?), and loops back to upstream tracing or same-level reframing if the check fails — until convergence. Applies across any domain — technical, process, organizational, decision-making, career. Triggers when the user uses evaluation or reconsideration language ("which is best", "root cause", "is this really needed", "sustainable long-term", "ほんとうに", "そもそも", "最善は", "根本は", "筋が良い"), when validating the soundness of someone's proposal from multiple perspectives, when the problem spans multiple layers (tech / process / org / regulation / UX / cost), or when clear trade-offs exist.

Review the local diff with flat-perspective sub-agents, fix findings in a loop, then create a PR via the create-pr skill once the review is clean

計画や設計について、ユーザーを徹底的に問い詰めて一緒に詰める。実装・着手の前に計画をストレステストしたいとき、または「詰めて」「グリルして」「問い詰めて」「ツッコんで」「設計レビューして」「ほんとうにこれでいい?」などのトリガー表現が出たときに必ず使う。

.claude/rules/ 配下のルールファイル (*.md) を新規作成・編集するときに、適切なメカニズム選択と paths スコープを保証する。ルールを書く前・書いた後に必ず適用する。

Orchestrate Epic issue as team leader. Delegate sub-issues to member agents in isolated worktrees. Members run in background and communicate via SendMessage.

Fetch a GitHub issue with its sub-issues, walk up to the parent Epic, list sibling issues under that Epic, fetch the body of each sibling issue, attach the implementing PR for each closed sibling with body and diff summary, gather related code, and present as a Japanese summary so the reader can grasp the whole task context including the actual contents of siblings and their PRs

Summarize GitHub Epic issue with sub-issues and related PRs. Use when reviewing epic progress or getting implementation overview.

Update PLANS.md content based on linked sub-issues status

Run the review-pr workflow on Codex instead of Claude — hand the PR number and the review-pr procedure to Codex via the codex:rescue subagent, then surface Codex's integrated review summary. Use when the user wants a PR reviewed by Codex / GPT, or asks to "review this PR with Codex".

Switch the current session into an existing git worktree (e.g. one created by /delegate-worker or `wt add`). Use when the user wants to enter, cd into, or follow a worktree — trigger on phrases like "enter the worktree", "worktree に入って", "ワーカーの worktree に移動して".

Create a new git worktree via `wt add` (firing .wt_hook.sh) and move the current session into it so the MAIN agent works there directly — no sub-agent. Use when the session started on main/master and the user wants to do the task themselves in an isolated worktree. Trigger on phrases like "worktree を作ってそこで作業して", "このセッションで worktree に入って直接やって", "work in a worktree yourself". For delegating to a background sub-agent instead, use /delegate-worker.

Delegate a task to a worktree-worker agent running in an isolated git worktree. Use this skill when you want a sub-agent to implement code changes, fix bugs, or investigate errors in a separate worktree — keeping the main session clean. Trigger on phrases like "delegate this to a worker", "fix this in a worktree", "have a worker handle this", or when the user wants code changes done in isolation.

Exit the current worktree session and return to the original working directory. Use when the user wants to leave a worktree entered via /enter-worktree or /delegate-worker — trigger on phrases like "exit the worktree", "worktree から出て", "元のディレクトリに戻って".

Fetch and execute mizchi's empirical-prompt-tuning skill at runtime. Use when evaluating or iteratively refining an agent-facing prompt (skill / slash command / task prompt / CLAUDE.md section / code-gen prompt) by dispatching an unbiased subagent, then improving until metrics plateau. Trigger right after creating or heavily revising such a prompt, or when agent misbehavior is suspected to stem from ambiguity in the instruction.

Comprehensive PR review with parallel multi-agent analysis, optional inline comment posting

Generate a HANDOVER.md file summarizing the current session's work. Use at the end of a session to preserve context for the next session.

Create a GitHub pull request using gh CLI, following the repository's own PR template and language conventions

Autonomously triage PR review comments and CI failures, apply necessary fixes, commit, push, and reply to comments. Designed for use with /loop 5m /babysit to continuously babysit PRs.

End-to-end workflow from GitHub issue to reviewed code. Reads issue, creates plan, implements code, and iterates review-fix cycles until quality passes.

Implement features from PLANS.md using the plan-driven-coder agent

End-to-end workflow that reads a GitHub issue, creates PLANS.md, and syncs it back to the issue comment. Combines read-issue, create-plan, and sync-plan into a single command.

Creates logical, well-structured git commits following Conventional Commits. Use when user mentions commit, committing, saving changes, or asks to organize changes into commits.

Updates the user's global Claude Code instructions file (~/.claude/CLAUDE.md, which is a symlink into the dotfiles repo). Use ONLY when the user explicitly asks to update, edit, add to, or remove from the global CLAUDE.md / global rules / global Claude instructions — typically while working in a different project where ~/.claude/CLAUDE.md isn't the primary editing target. Do NOT trigger for project-local CLAUDE.md files.

Triage and reply to PR review comments. Phase 1 categorizes comments by priority (must/investigate/info); Phase 2 posts threaded replies for resolved comments. Use when handling AI reviewer feedback (CodeRabbit, Copilot, etc.) or managing PR comments systematically.

Interactive coaching skill for polishing technical articles, tech blogs, and personal engineering essays. Starts from the user's rough draft and guides the article through meta-review, rewrites toward natural human-sounding prose, removal of clickbait/over-claims, title and description redesign, footnote curation, and commit granularity management. MUST trigger whenever the user mentions writing or editing a technical article in mdx/md, asks to review a draft, complains the writing feels AI-generated, asks for natural-sounding prose, wants to brainstorm titles or descriptions, wants meta-level feedback on structure/impact, prepares a post for Qiita/Zenn/Note/Medium/company blog, or otherwise wants a co-author and editor to refine an article together. Use proactively even when the user does not explicitly name "coach-tech-writing".

Create a sub-issue linked to a parent GitHub issue, honoring repository-specific issue conventions

Copies `claude -r <session-id> --fork-session` for the current session to the clipboard. Use when the user wants to fork the current session into a new branch from another terminal, or asks to copy the fork command, session id, or `/fork` command.

Commit changes, push to remote, and create a draft pull request

Commit changes and push to remote repository

Compress, reduce, optimize, or shrink documentation files. Use when docs are verbose, redundant, or need consolidation for clarity.

General guidelines for creating GitHub issues

Create a PLANS.md execution plan document for project management

Create a new Claude Code skill with best practices. Use when user wants to create a skill, add a SKILL.md, or extend Claude's capabilities.

Fetches a GitHub issue, implements the fix, and verifies it. Use when user wants to fix an issue, implement issue changes, or resolve a bug.

Creates a fixup commit and performs autosquash rebase

Merges a branch into current branch with conflict resolution. Use when user wants to merge, combine branches, or integrate changes.

Pulls latest changes from remote with conflict resolution. Use when user wants to pull, fetch changes, or update from remote.

Pushes commits to remote repository. Use when user wants to push, upload commits, or send changes to remote.

Rebases current branch onto target with conflict resolution. Use when user wants to rebase, update branch base, or linearize history.

Review local code changes against main branch