requirement-alignment

Requirement Intent Alignment

Overview

Ambiguity in a requirement is resolved through a funnel that gets cheaper per level, never by silent assumption. The MCP tools are the enforcement: they reject malformed questions, refuse to settle answers without a landing point, and gate delivery on mechanical lint.

The full law is the MCP prompt doc_analysis_playbook — read it in full before starting Step 0. This skill is the map; the playbook is the territory.

Dual-layer detection — the diagram is the instrument, not a deliverable. Layer 1 (reading): sweep the PRD text for explicit ambiguities (the eleven-category checklist). Layer 2 (drawing): do NOT wait for answers before drawing — hard-draw draft diagrams in the FIRST turn; every node/edge you cannot draw becomes a TBDn placeholder (state "???待确认" as TBDn / --> TBDn; never bare --> ???, lint's state-id charset eats it). Each placeholder is a gap: 🔧 technical → code evidence first, 📋 business → dispatch_question. Placeholders are cleared ONLY by code evidence or a human ruling — never by guessing. The first resolve_question is mechanically gated on the draft existing (analysis-draft.md with a mermaid block).

The Funnel

Resolve each gap at the cheapest level that can close it:

1. Code evidence (🔧 technical gaps first) — search the codebase (codegraph search_graph / trace_path, or grep). If the code has a single ground truth, inject it and settle with resolve_question(source="code"). Zero human cost.
2. Registered inference — no direct ground truth but a strong analogy? record_inference with an explicit evidence chain, batch-confirm with the human at session end via confirm_inferences. 🔴 Forbidden on core money flows and red-line rules — those need a human decision.
3. Group escalation (optional, sister intent-gate-service MCP) — judgment calls whose rightful answerer sits in a DingTalk group: intent-gate-service's group_dispatch with category 📋 business / 🔧 technical, @ the right role. One question in flight at a time. Core-flow questions carry your inference as recommended option 1, so the human nods instead of drafting. Skip this level when intent-gate-service is not mounted.
4. Dialog fallback (default) — the host asks the user directly, 3+1 options per question.