메뉴
Use this skill to delegate tasks to Claude Code CLI as a sub-agent. Triggers on: 'ask claude', 'run it with claude', 'claude -p', or when the user explicitly wants to spawn a separate Claude Code session for a task.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
Who yamadashy (Kazuki Yamada) is — a compact profile card to load when an agent needs to know the author's identity, background, or work. Triggers on: 'who is yamadashy', 'who am I', 'about the author/maintainer', 'introduce yamadashy', or when a task needs the user's bio, main products, or links.
Use this skill to delegate tasks to OpenAI Codex CLI as a sub-agent, including image generation/editing with GPT-Image-2. Triggers on: 'ask codex', 'run it with codex', 'codex exec', 'generate an image with codex', or when the user explicitly wants to use the Codex CLI for a task.
Propose the tasks on a repository that only a frontier-tier model can execute — the work the model one tier below or a routine engineer would fail at or do shallowly. The deliverable is a numbered pitch list ('I can do this here, and nothing cheaper can') for the human to greenlight. Triggers on: 'what could only you do in this repo', 'only-you', 'find frontier-worthy tasks', 'what work is worth handing to you', or when the user wants to know what to give the most capable model available.
Safely batch-merge Renovate's dependency-update PRs in the current repo. Lists open Renovate PRs and runs each through fixed safety gates — author is the Renovate bot with GitHub-verified (signed) commits, the update is non-major, the diff is confined to dependency files and clears a supply-chain scan, the new release is not brand-new (cooldown), CI is green, and the PR is mergeable — then auto-approves and merges only the PRs that clear every gate, holding the rest for manual review. Triggers on: 'merge the renovate PRs', 'merge dependency updates', 'handle renovate PRs', 'process renovate', 'update dependencies via renovate'.
日本語の文章を人間が読みやすいリズムに整えるスキル。黙読時に脳内で文を「声に乗せて」処理する仕組み（内在的韻律・サブボーカライゼーション）を前提に、音韻の紛らわしさ・韻律と統語の整合・文の単調さを点検して直す。日本語の文章を書く・推敲する・レビューするとき、あるいは「読みにくい」「リズムが悪い」「硬い」「読点の打ち方」「文末が単調」といった指摘や要望があったときは、明示的に「リズムを直して」と言われていなくても積極的に使うこと。コード・表・箇条書きが主体のドキュメントには使わない。
Vet a GitHub pull request and its author for supply-chain risk before reviewing or merging. Treats all PR/author text as untrusted to resist prompt-injection of the reviewing agent, investigates the author's account reputation, scans the PR diff for malicious patterns (obfuscation, install hooks, native-build/binding.gyp exec, credential harvesting, Trojan Source, hidden network egress, guard removal), and checks the change against current supply-chain attack techniques. Triggers on: 'vet this PR', 'vet this contributor', 'check this PR author', 'is this PR safe', 'supply chain risk', 'can we trust this PR', or before merging a fork PR from an unfamiliar author.
| name | claude-exec |
| description | Use this skill to delegate tasks to Claude Code CLI as a sub-agent. Triggers on: 'ask claude', 'run it with claude', 'claude -p', or when the user explicitly wants to spawn a separate Claude Code session for a task. |
A skill for invoking the Claude Code CLI as a sub-agent in non-interactive mode (-p / --print).
Run non-interactively with claude -p via the Bash tool.
claude -p --no-session-persistence --permission-mode default \
--disallowedTools Bash,Edit,Write -- "<task>"
For tasks that write files, explicitly allow only the tools you need via --allowedTools:
claude -p --no-session-persistence --permission-mode acceptEdits \
--allowedTools Read,Edit,Write -- "<task>"
| Flag | Purpose |
|---|---|
-p / --print | Run in non-interactive mode and exit |
--output-format json | Return the result as JSON (stream-json is also available) |
--model <name> | Select the model (opus / sonnet / haiku, etc.) |
--permission-mode <mode> | default / acceptEdits / plan / bypassPermissions |
--no-session-persistence | Do not persist the session to disk (requires -p) |
--allowedTools <list> | Restrict to a comma-separated allowlist of tools (e.g. Read,Edit) |
--disallowedTools <list> | Specify a comma-separated denylist of tools |
--append-system-prompt <text> | Append to the system prompt |
--add-dir <path> | Grant access to an additional directory |
--max-budget-usd <amount> | Set a cost ceiling |
--bare | Minimal mode (disables hooks, auto-memory, automatic CLAUDE.md loading, etc.) |
Pipe the prompt via stdin (the prompt argument can then be omitted):
echo "<long prompt>" | claude -p --no-session-persistence \
--permission-mode default --disallowedTools Bash,Edit,Write
claude -p --output-format json --no-session-persistence \
--permission-mode default --disallowedTools Bash,Edit,Write \
-- "<task>" | jq -r '.result'
The JSON includes result / total_cost_usd / session_id / duration_ms / modelUsage, among others.
--permission-mode bypassPermissions can be rejected by the auto-mode classifier (a safeguard against runaway sub-agents). For sub-agent use, prefer default + --disallowedTools or acceptEdits + --allowedTools first.-- before the prompt. Variadic flags such as --allowedTools / --disallowedTools also accept space-separated values, so without the separator they can swallow the following prompt argument as a flag value and produce an "Input must be provided..." error. A comma-separated list plus the -- separator is the safe form.timeout: 600000).