/plan — start a new planning round

This is plan space. Before any code, before any beads, before any agents. Resist the instinct to start implementing.

Slug: $ARGUMENTS (kebab-case; e.g. async-validation → .flywheel/plans/2026-04-20-async-validation.md).

Step 1 — Create the plan file

1. Today's date (UTC), then slug. Filename: .flywheel/plans/<YYYY-MM-DD>-$ARGUMENTS.md.
2. Copy the initial-plan prompt from .flywheel/prompts/plan-initial.md into a scratch area.
3. Fill in: intent, user workflows, constraints/invariants specific to this feature. Keep the Forge-wide invariants as-is (permission gate, LEGAL_TRANSITIONS, UI < 120 KB, etc.).

Step 2 — Send to 3–4 frontier models independently

Ideally: GPT Pro (Extended Reasoning), Claude Opus on the web, Gemini with Deep Think, and Grok Heavy. Each one solo, not combined.

Step 3 — Save each output

Save each model's response to a separate file:

.flywheel/plans/<date>-$ARGUMENTS-gpt-pro.md
.flywheel/plans/<date>-$ARGUMENTS-opus.md
.flywheel/plans/<date>-$ARGUMENTS-gemini.md
.flywheel/plans/<date>-$ARGUMENTS-grok.md

Step 4 — Synthesize

Run /plan-synthesize next. That skill feeds the competing plans into GPT Pro to produce a single hybrid.

Step 5 — Iterate

Run .flywheel/prompts/plan-refine.md in fresh GPT Pro conversations. Typically 4–5 rounds until suggestions get incremental. Plans routinely reach 3,000–6,000 lines — that's correct, not slop.

Step 6 — When to stop

Stop when:

• whole-workflow questions aren't moving around,
• architecture debates are settled,
• fresh models stop finding substantial missing features.

Then run /plan-to-beads.