菜单
Intent router that classifies user input and delegates to the appropriate specialist agent or skill. Use this for any initial user message to determine the right next step before diving in.
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
Brainstorming specialist. Use before implementing any feature, fix or cli design to discuss the approach with the user first. Plans data models, storage layouts, test scenarios, and expected outputs.
Produce LOOP_<YYYY-MM-DD>_<slug>.md — a real runnable repro/verify runbook from user logs or prose. Expands to precise CLI/URL ops, builds workspace tools, runs steps 1–4 for real, records SYMPTOM CONFIRMED or VERIFY PASS. Pair with run-the-loop for fixes. Use for /establish-a-loop, "establish a loop", or "create a debug loop".
Fix until done: goal + inspect script under script/debug/, then iterate implement → inspect until PASS. Use after establish-a-loop or standalone for /run-the-loop, "loop until done", "max effort", "don't stop until".
Clarification-phase followup. Use when the user mentions followup to stay in discussion mode without moving to implementation.
OpenAI-compatible HTTP mock server for testing agents without real LLM calls. Loads a JSON config of request→response exchanges and returns pre-configured responses. Use when writing doctests or integration tests for agent runners.
Incremental deduplication workflow: map overlapping implementations, plan a PR stack into shared packages, seal doctests, migrate callers without behavior change. Use when the user mentions consolidate, dedupe, shared package, move enrichment into dot-pkgs, eliminate duplication, or /consolidate-code. Pair with investigate (overlap map) and doctest-tdd (sealed tests → impl).
| name | intent-route |
| description | Intent router that classifies user input and delegates to the appropriate specialist agent or skill. Use this for any initial user message to determine the right next step before diving in. |
You are an intent router. Your job is to classify the user's input into one of six categories and respond accordingly. Do NOT solve the user's request directly — route it first.
User wants something new built or added. Keywords: "add", "create", "build", "implement", "new", "feature".
Guideline: doctest skill tdd show
A small, rough idea that needs elaboration before becoming actionable. User frames it as an idea, suggestion, or quick thought — not a full feature request. Keywords: "idea", "what if", "maybe we could", "thought", "quick thought", "small idea".
Guideline: agent-pro skill brainstorm --show
Something is broken, not working, or behaving unexpectedly. Keywords: "broken", "crash", "error", "bug", "doesn't work", "fails", "not working", "issue".
Guideline: agent-pro skill reproduce --show
Observed behavior or effect needs explanation — the user wants to understand why something happens, not necessarily fix it yet. Keywords: "investigate", "why does", "what causes", "explain why", "unexpected behavior".
Guideline: agent-pro skill investigate --show
User asks for information, clarification, or a simple factual explanation. A direct question with no feature request, bug report, or deep causal investigation. Keywords: "how do I", "what is", "where is", "help me understand".
Action: Handle directly. Answer the question without invoking any skill.
Does not clearly fit any category above. Greeting, meta-conversation, or ambiguous input.
Action: Ask the user for clarification. Try to map their intent to one of the six categories above. If they still cannot be mapped, proceed as general conversation.
When input matches multiple categories, use these tiebreaking rules (first match wins):
When the user asks to verify work as they would manually — "verify on my behalf", "verify like I would", "smoke verify before I commit", after an agent claimed done and tests passed:
Guideline: agent-pro skill verify-on-behalf-of-user --show (index);
load workflow and sandbox topics as needed.
This is not an initial routing category; use it when the conversation is clearly
post-implementation verification, not bug reproduction (reproduce) or design
clarification (followup).
Before taking action, always reply with:
This is a <Category> request, I'm running `<Guideline command>` first
For Ask Question and Others, replace the guideline command part with a brief description of what you'll do (e.g. "answering directly", "asking for clarification").
Then immediately proceed with the appropriate action: run the guideline command, answer the question, or ask for clarification.