The workflow is simple: write specs with AI in chat, then /plan to build.

Specs are just structured conversations — charter, PRD, ADD, WBS, roadmap — saved to files. There's no special command. /plan is the real Claude Code feature: it reads your specs, proposes an implementation approach, and gates code behind your approval.

New Project

Spec first, code second. In one focused day:

1. Charter — problem, audience, scope (30 min)
2. PRD — requirements split P0/P1/P2 (2–4 hrs)
3. ADD — architecture, ADRs, tech choices (2–3 hrs)
4. WBS + Roadmap — bite-sized work packages (1–2 hrs)
5. Set up CLAUDE.md — point to specs, set current milestone

Then for each work package: /plan → review → approve → implement.

Mid-Project

Extract specs from decisions you already made, don't invent new ones.

1. Update CLAUDE.md now — what works, what's broken, what's next
2. Charter backwards — describe what you built and what you won't add
3. PRD from your head — broken = P0, planned = P1, dreaming = P2
4. ADD from your code — document patterns you actually use, not aspirational ones
5. WBS from the gap — diff PRD against code; the gap is your work packages

Then same loop: /plan per work package.