메뉴
Delegate a task to Cursor's CLI agent (code review, Q&A, planning) for a second opinion from a non-Claude model
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
Analyze recent Claude Code sessions and report where the most tokens (and estimated cost) were burned, with data-driven tips to cut usage and avoid rate limits. Scans ~/.claude/projects session logs over a window (default last 5h, configurable), ranks the heaviest sessions and projects, breaks down cache efficiency and model spend, and renders a minimalist report — in a cmux markdown panel when cmux is available, otherwise as an ASCII summary in the terminal. Use this whenever the user wants to know where their tokens went, why they're hitting (or nearing) usage/rate limits, what their recent Claude Code activity cost, which sessions or projects are the most expensive, or how to optimize token usage. Trigger on phrases like "token burn", "where did my tokens go", "token/usage report", "what's burning tokens", "session cost", "how much have I spent on Claude", "why am I hitting rate limits", "analyze my claude sessions", "token breakdown", "/token-burn", and "/burn".
Packages the current chat (or one or more specific issues/topics) into self-contained handoff files and launches each in its own fresh `claude` session with a dedicated, meaningfully named git worktree — so work continues without manual context copy-paste. Each child claude inherits the CALLER session's permission mode (no hardcoded bypass). Requires cmux: by default each handoff opens as a new cmux workspace; with `--tabs` (or "in this/the same workspace", "as tabs") they open as tabs in the current workspace. Use when the user wants to hand off work to a fresh Claude in its own worktree instead of continuing inline. Trigger phrases: "handoff to worktree", "handoff-to-worktree", "spin this off into its own claude", "fork this into a worktree", "run it in a separate claude". No argument = handoff the whole chat (one worktree); argument(s) = one handoff/worktree per issue or topic.
Hands-free voice loop for driving Claude by voice. Combines the say skill (Gemini TTS audio out), a focus-pull back to the caller's cmux prompt, Wispr Flow dictation (voice in), and an appendix-stop listener so the user ENDS and SUBMITS each turn by saying the word appendix. When active, every reply is spoken aloud and the mic is auto-armed, so the user answers and sends without touching the keyboard. Use when the user asks for hands-free mode, to control Claude by voice, to start the voice flow, or voice mode on. Turn it off with hands-free off, stop voice mode, or stop the voice flow.
Summarize and simplify the last agent message, then speak it aloud via Gemini TTS (Vertex AI, voice Charon), in whatever language the user is using in the session. Falls back to sag/macOS say.
Align recently written or changed code so it matches the surrounding project's established conventions — formatting, naming, imports, comments, and idioms — learned from the codebase itself, then apply the fixes. Use this whenever the user wants their changes to "match the codebase / our conventions / the rest of the app", asks to make a diff or PR style-consistent, wants a style pass or style review before opening a PR, says code "feels off" or inconsistent with the project, or invokes /code-style. Works in any language or framework — it discovers the rules from the repo rather than assuming them. Reach for this even when the user doesn't say the word "style" but clearly wants new code to blend in with existing code — including trimming AI-generated over-commenting down to the codebase's own comment density.
Two-phase research workflow for Medium articles. Phase 1 discovers candidates via RSS tag feeds plus DuckDuckGo content search; Phase 2 spawns one subagent per selected article to fetch full content through the Freedium mirror (paywall bypass) and return a structured digest. Use whenever the user wants to research a topic on Medium, surface trending or recent articles, pull full text from member-only Medium posts, or build a reading list. Triggers include "/medium-research <topic>", "research medium <topic>", "medium research", "find medium articles about X", "what's trending on medium about X". The user has paid Medium membership and treats Freedium use as acceptable for personal research. Do not invoke for non-Medium research (use /last30days for cross-platform).
| name | cursor-agent |
| description | Delegate a task to Cursor's CLI agent (code review, Q&A, planning) for a second opinion from a non-Claude model |
| allowed-tools | ["Bash(cursor-agent:*)","Bash(git diff:*)","Bash(git log:*)","Bash(git status:*)"] |
| when_to_use | Use when the user wants to run cursor-agent (Cursor's CLI). Triggers: 'cursor review', 'ask cursor', 'use cursor', 'review with cursor', 'second opinion from cursor', 'cursor-agent', 'use cursor-agent'. Also use for verifying code/plans/texts/UI with a different model than Claude. |
| argument-hint | [task for cursor-agent, e.g. 'review this branch'] |
| arguments | ["task"] |
Delegate a task to Cursor's headless CLI agent (cursor-agent). Use this to get a second opinion from a non-Claude model on code reviews, plan critiques, text/copy review, or general Q&A.
$task: The task to hand to cursor-agent. Usually a review or question prompt. If the user adds a model hint (e.g. "with Opus", "using GPT", "pomocí Sonnetu"), extract it and pass via --model.Run cursor-agent headless, surface its output to the user, and add a brief Claude TL;DR so the user doesn't have to re-read a wall of text.
Extract from $task and the surrounding conversation:
--model claude-opus-4-7-medium--model claude-4.6-sonnet-medium--model gpt-5.3-codex--model gemini-3.1-pro--model auto--model (cursor uses its default: composer-2-fast)--mode plan--mode (cursor can write)--mode ask (read-only, default)Success criteria: Flags decided, prompt ready.
If the task is a code review:
main (git diff origin/main...HEAD), or uncommitted changes (git diff HEAD)."Review the branch
<name>against main. Usegit diff origin/main...HEADto see the actual changes. Focus on real issues (correctness, edge cases, security); skip nitpicks. Report under ~400 words withfile:linereferences."
For non-review tasks (ask/plan/text/UI), pass the task as-is.
Success criteria: Final prompt string ready for the CLI.
Invoke headless via Bash:
cursor-agent -p --output-format text [--mode <mode>] [--model <model>] "<prompt>"
-p (print/headless).--output-format text unless the task needs structured output.--workspace unless the user explicitly asked for a different dir.--mode ask or --mode plan), no extra permission flags are needed. For headless write mode, do not silently add --force or --yolo; ask the user first.tail -n 200 if you suspect a very long response and want to cap context use. By default, take the full output.Success criteria: cursor-agent returns (exit 0) with a response. If it errors (auth, model not available), surface the error to the user instead of hiding it.
Present to the user in this order:
Keep the summary tight — the goal is to save the user from re-reading, not to replace cursor's output.
Success criteria: User sees both cursor's raw output and a Claude-curated action list.
cursor-agent --list-models shows every available model ID. If the user asks for something exotic (e.g. "thinking tier", "max tier"), check the list first.cursor-agent reads/writes inside the current workspace like any CLI tool — it is not sandboxed by default. Treat its output the same way you'd treat a teammate's PR comment: trust but verify.cursor-agent interactively themselves (no -p).