Designit

Starts from a single high-level plan and progressively zooms in, one round at a time, until the details are locked. No coding — the output is a decision, not an implementation.

When to use this

• Before building anything whose shape isn't already obvious.
• Skip for a change you could describe fully in one sentence — go straight to the edit, or to /plan-doc if it still deserves a record.

Procedure

1. Read .specify/memory/constitution.md.

2. State your assumptions upfront, and ask clarifying questions rather than guessing at anything that would change the design. Read the codebase to ground them — do not guess the tech stack.

3. Draft the design at one level of detail, then present it and stop. This is a human gate (interactive — answered in-session per the constitution's gate protocol). The user decides: redo / confirm and continue (zoom in one level) / confirm and finalize.

4. Repeat until every Critical/Major decision is settled or the user says "confirm and finalize". Each round adds real resolution — structure, boundaries, libraries, the decisions and their reasons — never a restatement of the last.

5. On finalize, write the spec to specs/NNN-<slug>/spec.md from .specify/templates/spec-template.md, status: accepted. Take NNN from the specs/ index. Then hand off: /plan-doc reads this spec and writes the plan.

The spec must carry

Produce it in the template's shape so the write is a copy, not a rewrite:

• Problem — why the work exists, what is currently true, the constraint that forced the design.
• Considered Approaches and the Decision with its reason — the choices that were actually contested.
• Design — enough detail to plan against without re-deriving it.
• Non-Goals — what was deliberately left out, so it isn't re-litigated mid-build.

What not to do

• Don't write code, and don't edit the target files "to check" — read only.
• Don't advance a round without the user's explicit word; "looks good" on one level isn't approval of the next.
• Don't leave a decision whose consequences you haven't drilled into.
• Don't break a fuzzy question into neat sub-questions. If you can't state it in one clear sentence yet, name the area and move on — it sharpens once earlier decisions land. The test is whether you can ask it precisely, not whether you can answer it.