Create SDD PRD

1. Extract the problem, outcome, and feature name from the request. Resolve tasks/prd-[slug]/prd.md; a new slug is [NN]-[name], where [NN] is the next two-digit number after the existing folders, in implementation order, and [name] is kebab-case. If the PRD exists, reuse it without overwriting. Update only when authorized; preserve unchanged IDs. When tasks/prd-[slug]/context-snapshot.md exists, load it through the Load branch of .agents/skills/sdd-snapshot/SKILL.md before reading sources. Output: unambiguous destination and operation; request only missing information that prevents identifying them.
2. Inventory users, journeys, metrics, FR, NFR, constraints, dependencies, accessibility, and out-of-scope items. Consult local evidence, including docs/specs/, before researching public rules or integrations in primary sources; other features' artifacts in tasks/prd-* are not evidence of current behavior, which lives in the code, and are not edited here; send read-only explorers only when the local evidence spans many files. Record origin and distinguish fact, assumption, and product decision; ask only about gaps that change scope or acceptance. Output: every obligation has evidence or an explicit assumption; blocking decisions are identified.
3. When drafting, read assets/prd.template.md in full. Use stable IDs OBJ-01, US-01, FR-01, and NFR-01, each with observable acceptance. Keep architecture and sequencing in the TechSpec. Record supplied stack and platform constraints. Output: all applicable sections filled; unique IDs; no requirement without acceptance.
4. Write only the PRD and report path, pending items, and obligations added, changed, or removed. Point out derived artifacts to revalidate without editing them. In the orchestrated flow, return the artifact for product HIL; previous authorization applies to the scope it covers. In standalone use, run the session pause from that reference with sdd-create-techspec as next step. Output: reviewable PRD on disk; source divergences and conflicting IDs resolved or explicitly pending.

Read each source once per version; use links for extensive evidence. On updates, preserve unchanged sections to reduce diff and context.