메뉴
Drive Chrome from the shell to browse, click, fill forms, run JS, and screenshot. Use for dynamic sites, web UIs, or content curl cannot reach.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
Enforces staged execution discipline on large tasks: a written stage plan, parallel delegation where the runtime supports it, a failable verification check at each stage, and a skeptical self-review before delivery. Trigger when the user explicitly asks ("do this thoroughly", "be systematic", "deep work mode") OR when the task objectively spans multiple files, multiple sources, or multiple sessions. Do NOT trigger on ordinary multi-step requests that a direct attempt handles fine. For a run pinned to a specific model, use fable-sonnet or fable-haiku instead.
Structured brainstorming when the user asks to explore options, compare approaches, or solve an ambiguous product, prompt, architecture, workflow, or design problem. Generates diverse options, clusters ideas, scores tradeoffs, and converges without generic ideation bloat. Avoid for straightforward implementation/debugging tasks.
Design lightweight evaluations for prompts, skills, agents, and workflow changes: before/after test prompts, success criteria, negative cases, edge cases, regression risks, and manual or scripted eval plans.
Review, debug, and improve user-provided or repo-visible prompts and agent instructions: AGENTS.md, skills, custom instructions, model configs, and prompt libraries. Optimizes for contradiction removal, trigger quality, context bloat reduction, eval ideas, and minimal high-leverage edits.
Search and read Atlassian Confluence content.
Interact with GitLab using `glab` CLI.
SOC 직업 분류 기준
| name | web-browser |
| description | Drive Chrome from the shell to browse, click, fill forms, run JS, and screenshot. Use for dynamic sites, web UIs, or content curl cannot reach. |
Uses agent-browser — Rust CLI over CDP. Daemon persists Chrome between calls.
--session <session-slug> isolates from the user's browser; pick a task-specific kebab-case slug (e.g. grafana-improvement, jira-triage). Use batch to run multiple steps in one invocation; fall back to single calls only when you need to inspect output between steps (snapshot → reason → click).
agent-browser batch --session <session-slug> \
"open https://example.com" \
"snapshot -i" \
"screenshot out.png" \
"close"
Refs @eN come from the latest snapshot. Traditional CSS selectors also work (click "#submit").
Headless by default. Add --headed to show the window. It only takes effect when the daemon starts — if you see ⚠ --headed ignored: daemon already running, run agent-browser --session <session-slug> close and retry.
snapshot -i — compact accessibility tree, best for agent reasoningget text @e1 / get html @e1 / get value @e1 / get title / get urleval '<js>' (--stdin for piped JS)find role button click --name "Submit"
find text "Sign In" click
find label "Email" fill "me@example.com"
is visible|enabled|checked <sel> — exit code reflects state.
agent-browser --help and agent-browser <cmd> --help for the full list (waits, tabs, uploads, pdf, recording, sessions). Prefer one-shot commands composed with shell over large eval blobs.