Planning — Brainstorm + Write Plans

TL;DR

• Use before any feature, behavior change, or 3+ step task
• Phase A (Brainstorm): clarify intent → 2-3 options → recommend → scope
• Phase B (Write): emit openspec/changes/<name>/{design.md,tasks.md}
• Handoff out: .cm/handoff/plan.json (goal, decisions, first 3 tasks)
• Next: cm-tdd or cm-execution

When to Use

• Creating features, components, or functionality
• Modifying behavior; multi-step tasks; user-visible change
• Skip only for trivial single-line edits or pure questions

Full Protocol

Phase A: Brainstorm

1. Intent — Ask clarifying questions. Don't assume scope. Surface hidden requirements.
2. Options — List 2-3 approaches with pros/cons. Recommend one with reasoning.
3. Scope — Must-have vs nice-to-have; edges to handle vs explicitly skip.
4. Design — Data flow, component boundaries, API contracts. UI work → cm-design-system.

Red flags (STOP): code before brainstorm; assuming intent; skipping scope; "it's simple."

Phase B: Write Plan (OpenSpec format)

Write to openspec/changes/<initiative-name>/:

design.md

# Design: [Goal]
## Context & Technical Approach
## Proposed Changes
### [Component/File]
## Verification

tasks.md

# Implementation Checklist
- [ ] 1.1 ...
- [ ] 1.2 ...
- [ ] Verification testing

Plan rules: small testable steps (15-30 min each), order by dependency, verification per step. No vague steps.

Step FINAL: Emit Handoff + Update Continuity

Write .cm/handoff/plan.json:

{
  "schema": "plan@1",
  "goal": "...",
  "decisions": ["..."],
  "first_tasks": ["1.1", "1.2", "1.3"],
  "openspec_path": "openspec/changes/<name>/"
}

Update .cm/CONTINUITY.md:

• Active Goal → plan goal
• Next Actions → first 3 tasks
• Current Phase → "planning"
• Working Context → key decisions

Token win: next session reads CONTINUITY (~200 tok) instead of full plan (2000+).

Integration

Karpathy Discipline — Think Before Coding

Before writing the plan, surface what's uncertain instead of guessing:

• State assumptions explicitly. List them in the plan; mark each as "verified" or "needs confirmation".
• Multiple interpretations? Present them side-by-side with tradeoffs. Don't pick silently.
• Simpler path exists? Say so. Push back on over-scoped requests.
• Confused? Stop. Name the confusion. Ask. Cheaper than building the wrong thing.

Self-test: "Could a reviewer point to anything in this plan I'm guessing about?" If yes, mark it.

Anti-Patterns

• ❌ Coding before plan exists
• ❌ Skipping handoff JSON (downstream skills lose context)
• ❌ Vague tasks ("refactor the code")
• ❌ Hidden assumptions — pick one interpretation without telling the user

The Bottom Line

Think before you build. Document before you code. Emit handoff so the next skill picks up cold.