andrem-sec
andrem-sec/psc-comet
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。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/andrem-sec/psc-comet --skill intent-router命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
基于 SOC 职业分类
正在显示 SKILL.md
| name | intent-router |
| description | Classify ambiguous requests and route to the correct skill or agent before any execution begins |
| version | 0.1.0 |
| level | 1 |
| triggers | ["not sure where to start","what should I do with this","how should I approach","where do I begin","ambiguous request before starting work"] |
| context_files | ["context/project.md"] |
| steps | [{"name":"Restate","description":"Restate the request in one sentence to confirm understanding. If restating reveals ambiguity, ask one clarifying question before classifying."},{"name":"Classify","description":"Assign exactly one category from the routing table. If it could be two categories, pick the one that must happen first."},{"name":"Route","description":"Name the first skill or agent to invoke and state why it is first — not what the full pipeline looks like."}] |
Classify before executing. Ambiguous requests that go straight to implementation end up solving the wrong problem. One classification step costs almost nothing; backtracking from the wrong path costs a lot.
Without intent routing, Claude defaults to the most familiar interpretation of an ambiguous request and starts implementing. A request like "fix the auth flow" could be a bug fix, a security review, a refactor, or a new feature — each requiring a different starting skill. Picking wrong means work that has to be undone.
| Category | Signals | First Skill / Agent |
|---|---|---|
| New feature | "add", "build", "implement", "I want", new capability | deep-interview → prd → feature-pipeline |
| Bug fix | "broken", "not working", "error", "failing", regression | debug-session → fix-pipeline |
| Refactor | "clean up", "reorganize", "too complex", no behavior change | refactor |
| Security review | "is this secure", "check for vulns", reviewing local code | security-gate |
| Security operation | scanning, testing external systems, autonomous probing | roe → docker-sandbox |
| Architectural decision | "how should we structure", multiple valid approaches, high impact | consensus-plan |
| Code review | "review this", "look at my changes", pre-merge check | code-review |
| Deploy / release | "deploy", "ship", "release", "push to prod" | security-gate → deploy |
| Unclear | Cannot classify with confidence | deep-interview |
One category only. If a request spans multiple categories, pick the one that must happen first. A "secure new feature" is a new feature — security review comes after implementation, not before.
"Unclear" is a valid classification. Do not force a category to avoid the overhead of a clarifying question. Wrong category costs more than one question.
Security operations require ROE before anything else. Any request that involves scanning, probing, or testing systems — even systems the user owns — routes to roe first. No exceptions.
Ambiguity signals:
Intent: [category]
Reason: [one sentence — why this category over alternatives]
First step: [skill or agent name] — [what it will accomplish]
If unclear:
Intent: unclear
Ambiguity: [what is missing]
Clarifying question: [one question that resolves the ambiguity]
Do not present the full pipeline at this stage. Route to the first step only. The pipeline emerges from the work.
Do not reclassify after the user confirms. If the classification was wrong, that is new information — run intent-router again with the corrected understanding.
Do not skip routing because the request "seems obvious." The obvious interpretation is often not the intended one.
roe before any execution