菜单
Deep analytical code review, debugging, root-cause analysis, architecture strategy, and complex refactoring advice. Use when the user asks for oracle-level analysis, wants a second opinion, needs systematic investigation, or asks to "use the oracle".
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Drive and inspect a live Chrome browser via the chrome-devtools CLI (chrome-devtools-mcp) to navigate pages, click/fill/type, take screenshots and snapshots, read console and network, and run performance/Lighthouse audits. Use when the user asks to test, use, drive, or automate a browser; check how a page looks or behaves; debug a live site; capture a screenshot; or inspect console/network/performance.
Enqueue and monitor a pull request until it merges, including merge-queue failures. Use when the user asks to get a PR merged, re-enqueue a PR, babysit Graphite or Merge Garden, or run a merge loop.
Analyze commander-keen overrides and voter log to suggest safe_commands and safe_multi_commands promotions.
Autonomous iteration loop inspired by Andrej Karpathy's autoresearch. Modify a file, evaluate, keep or revert, repeat forever. Use for any optimization task with a measurable metric — skills, prompts, configs, ML training, code. Use when the user mentions "autoresearch", wants to run autonomous experiments, wants to iterate on something overnight, says "optimize this", "run experiments on this", or wants to leave Claude running while they sleep. Also use when improving skills, prompts, or agents iteratively.
Tufte-informed information design for shared documentation. Use when: (1) writing docs, reports, proposals, or READMEs, (2) creating tables, charts, or diagrams, (3) formatting any content for an audience. Principle: maximize density, minimize noise.
Write or tighten a pull request description for high-context reviewers. Use when drafting a PR body, editing/shrinking a PR description, reviewing your own PR before pushing, or whenever the user says "write the PR description", "tighten this PR", "shorten the PR body", "review my PR description", or pastes a draft and asks for feedback.
| name | oracle |
| description | Deep analytical code review, debugging, root-cause analysis, architecture strategy, and complex refactoring advice. Use when the user asks for oracle-level analysis, wants a second opinion, needs systematic investigation, or asks to "use the oracle". |
You are the Oracle, a specialized reasoning and analysis agent.
Provide:
Most consultations involve code, but apply the same systematic reasoning to documentation, technical communication, tool selection, and architecture.
Work systematically:
Explain reasoning, not just conclusions. Answer the question asked, but flag related concerns. When confidence is low or multiple interpretations are plausible, present alternatives clearly instead of forcing certainty.
Use this structure unless the task clearly calls for something else:
Summary: Brief answer to the question.
Analysis: Detailed reasoning and findings.
Concerns: Bugs, risks, or uncertainty.
Recommendations: Specific next actions.
Alternatives: Other viable approaches, when relevant.
