メニュー
CLI syntax for dispatching prompts to external model agents (claude, cursor, codex, gemini). Use when spawning a second-opinion review, running prompts through alternate models, or scripting multi-model workflows.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
Run end-to-end GitHub issue triage using the local triage scripts and parallel subagent scoring. Use when asked to create, refresh, or rerun issue triage markdown with complexity, difficulty, benefit, scope, and execution-lane recommendations (`agent-implementation`, `agent-assist-human`, `human-only`).
Find and resume a past Claude Code, Codex, or Cursor Agent conversation by searching session history files by keyword. Use this skill whenever the user wants to find a previous conversation, can't remember which directory or agent they were using, wants to look up work from a past session, needs to resume a session they can't locate with /resume, codex resume, or agent --resume, or says something like "I know we discussed X somewhere" or "where was that session about Y". Infers search scope, agent, and keywords from the user's description and presents matching sessions with summaries before asking which to resume. Always use this skill rather than manually searching ~/.claude/, ~/.codex/, or ~/.cursor/ yourself.
Use for commit-level Git workflows in non-`.jj/` repositories: stage one logical change, write or revise commit messages, amend, create fixup commits, split bundled edits, curate history with rebase or autosquash, and verify bisectability. In repositories containing `.jj/`, use this skill only for explicit Git-only commit intent; generic commit intent routes to `jujutsu` plus `jj-commits`. Trigger on requests such as "commit this", "check it in", "ship it", "stage and commit", "amend", "fixup", "clean up history", "reword/split commits", or "draft/rewrite the commit message". Also use for indirect commit-message creation when drafting plans that specify commits (including `superpowers:writing-plans` outputs), and when writing commit message text in design docs, PR notes, or other artifacts.
Guidelines for designing and implementing command-line interfaces. Use when creating a new CLI tool, adding subcommands or flags to an existing one, implementing help text, designing error messages, or structuring CLI output. Also use when reviewing CLI code for usability, when the user asks about CLI best practices, or when generating argument parsers, config handling, or shell completions. Applies to any language.
Apply Christopher Larson's preferred Bash and POSIX shell script style when creating or substantially editing user-facing shell scripts. Use when writing or modifying shell utilities under `scripts/` or `script/` (including extensionless shebang executables, pre-commit entry scripts, and CI helpers), `.sh` files, shell snippets intended for the user to keep, or any Bash or POSIX sh program that will become part of the maintained codebase. Do not use for temporary test scripts, disposable intermediate artifacts, or scripts written solely for AI-agent execution.
Use for generic commit-related intent in `.jj/` repositories. Must be paired with `jujutsu` in the same turn; never use it as a standalone skill. Explicit Git-only `.jj` commit intent routes to `git-commits`.
| name | dispatch-external-model |
| description | CLI syntax for dispatching prompts to external model agents (claude, cursor, codex, gemini). Use when spawning a second-opinion review, running prompts through alternate models, or scripting multi-model workflows. |
CLI commands for running prompts through external model agents non-interactively.
| Tool | Command |
|---|---|
| Claude | claude -p "prompt" |
| Cursor | agent -p --trust "prompt" |
| Codex | codex exec "prompt" |
| Gemini | gemini -p "prompt" |
claude -p "prompt"
-p — print mode (non-interactive, output to stdout)--model <model> — specify model (e.g., sonnet, opus)settings.json permission allowlists by defaultAuto-safe (on request): Add --permission-mode acceptEdits to auto-approve file edits, or --permission-mode auto for broader auto-approval within configured limits.
Full YOLO (on request + user approval): Add --dangerously-skip-permissions to bypass all permission checks. Only use when the user explicitly requests this level of trust.
agent -p --trust "prompt"
-p — print mode (non-interactive, output to stdout)--trust — trust the current workspace (required for headless execution)--model <model> — specify model (e.g., sonnet-4, gpt-5)Sandbox requirement: Cursor requires macOS system certificate trust settings. Run with dangerouslyDisableSandbox: true in Claude Code.
Auto-approve (on request): Add --yolo or --force to auto-approve all tool calls. Requires user approval before invoking.
codex exec "prompt"
-m <model> — specify model-o <file> — write last message to file--json — output as JSONLSandbox requirement: Codex requires write access to ~/.codex/sessions. Run with dangerouslyDisableSandbox: true in Claude Code.
Auto-safe (on request): Add --full-auto for sandboxed automatic execution (auto-approves within workspace-write sandbox).
Full YOLO (on request + user approval): Add --dangerously-bypass-approvals-and-sandbox to skip all prompts and sandboxing. Only use when the user explicitly requests this level of trust.
gemini -p "prompt"
-p / --prompt — non-interactive mode-o json — output as JSON-o stream-json — output as JSONLSandbox status: Unknown — not yet tested. May require dangerouslyDisableSandbox: true.
Auto-safe (on request): Add --approval-mode auto_edit to auto-approve file edits only.
Full YOLO (on request + user approval): Add --approval-mode yolo to auto-approve all actions. Only use when the user explicitly requests this level of trust.
When dispatching from within a sandboxed agent (e.g., Claude Code), the sandbox applies to child processes. This means all external model CLIs require sandbox escalation, even Claude calling Claude.
From Claude Code: Use dangerouslyDisableSandbox: true in the Bash tool for all external model dispatch. The child process inherits sandbox restrictions that block session/state directory writes.
From Cursor: Cursor has no per-tool-call sandbox escape (unlike Claude's dangerouslyDisableSandbox). Options:
permissions.json (terminalAllowlist)--yolo flag — requires explicit user request and approvalFrom Codex: Codex's --full-auto runs in a workspace-write sandbox. External CLIs that need to write outside the workspace will fail.
When dispatching to external models, use the minimum trust level needed:
Default — Use the base commands above. The external agent may prompt for approvals (which fails non-interactively for read-only tasks, but is safest).
Auto-safe — Use --permission-mode acceptEdits (Claude), --full-auto (Codex), or --approval-mode auto_edit (Gemini) when the calling agent explicitly requests autonomous execution. These maintain sandboxing or limit auto-approval to edits.
Full YOLO — Use --dangerously-skip-permissions (Claude), --dangerously-bypass-approvals-and-sandbox (Codex), --approval-mode yolo (Gemini), or --yolo (Cursor) only when: