메뉴
Run interactive CLI commands (Metro, Vite, dev servers) with PTY capture and agent-friendly log tail. Use when the user wants to run a long-lived process with log capture, grep session logs, follow live output, or inspect session state.
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
| name | task |
| description | Run interactive CLI commands (Metro, Vite, dev servers) with PTY capture and agent-friendly log tail. Use when the user wants to run a long-lived process with log capture, grep session logs, follow live output, or inspect session state. |
tools taskPTY-aware command wrapper with ordered JSONL capture.
tools task get --session <name> — first command; shows state, files, flags cheat sheettools task logs --session <name> --raw | grep PAT — grep-safe log readtools task tail --session <name> --follow — live followget → stderr only (don't pipe)logs/tail → use --raw or --jsonl for piping stdout--follow, not -f)~/.genesis-tools/task/sessions/<name>.jsonl| Command | Purpose |
|---|---|
task run --session NAME -- CMD | Foreground run + capture |
task get --session NAME | Session info panel |
task logs --session NAME | Read logs (add --tail --follow for live) |
task tail --session NAME --follow | Live follow |
task sessions | List sessions |
task clean --session NAME | Remove session files |
task dashboard open | Open unified dashboard |
tools task run --session metro -- npx react-native start
tools task get --session metro
tools task logs --session metro --lines 100 --raw | grep BUNDLE
tools task tail --session metro --follow
See tools task --readme for full documentation.
Control macOS UI via the Accessibility API and record short screen captures reviewed frame-by-frame. Use when automating native apps — clicking buttons, filling forms, reading element state, finding elements, getting window bounds — or when the user wants a short screen recording of an animation/transition/flicker reviewed via a contact sheet, optionally pushed to a vitrinka board. Triggers on "click button in app", "fill form in native app", "automate macOS app", "list UI elements", "find button", "read text field value", "get window position", "interact with native app", "record the screen", "capture this transition", "Xs recording", "Nfps", "watch this animation", "why does it jump/flicker", "record this app for a few seconds", "push the recording to a board".
Drive a real, extension-loaded Chrome/Brave browser for the GenesisTools YouTube extension via chrome-devtools-mcp as a standalone MCP client — no Claude Code config edits, no restart. Use whenever the user wants to test, debug, screenshot, or click through the YouTube extension's UI (side panel tabs, popup, content script), asks to "load the extension and click X", wants a screenshot of the extension mid-interaction, or needs to find exact pixel coordinates to click something inside a web page (not a native macOS app — for that, use the screen-capture skill instead). Also use this whenever chrome-devtools-mcp's own tools are attempted and fail with "no browser" or launch a blank vanilla browser with no extension loaded — that means the server wasn't wired to a real browser, and this skill is the fix.
Cross-agent messaging protocol via `tools agents`. Use when the main agent is about to spawn subagents that should be able to message each other (or back to the main agent) during their work. Teaches the login (auto-registers) / message / discover / listen pattern, the stream-vs-once receive modes, and the no-loss delivery contract.
Empirically evaluate and trim ANY agent-instruction file — user/project CLAUDE.md, CLAUDE.local.md, AND AGENTS.md (the cross-tool standard read by Codex, Cursor, Gemini CLI) — to find rules a clean model already knows, rules the repo/hooks already enforce, and internal duplication, then propose evidence-backed cuts. Autodetects which files exist and which the harness actually loads, runs headless claude with ALL instruction files EXCLUDED (verified by canary) against MULTIPLE models, and classifies every rule KEEP/SHRINK/CUT with a multi-provider rubric. Use when the user says "improve agents.md", "evaluate agents.md", "trim CLAUDE.md", "improve claude md", "is my CLAUDE.md bloated", or "/improve-agents-md [path]".
Create and maintain a compaction-proof progress file (.claude/plans/<plan>.handoff.md) next to an implementation plan, so any executor - fresh session, post-compact resume, or parallel subagent - knows exactly where the work stands and which plan lines to read next. Invoke when starting to execute a plan, when resuming one, when handing work to another agent, or when the user says "create a handoff" / "prepare the handoff".
Write comprehensive implementation plans (subsumes writing-plans - bite-sized TDD tasks, exact paths, complete code) hardened with an executor-proof contract - all judgment spent at plan time, REASON/VERIFY/ON-FAIL per task, no-improvisation covenant - so even the weakest model executes to strong-model-quality output. Invoke INSTEAD of writing-plans whenever the plan may be executed by another session, subagent, or weaker model.