Heavy-prompt builder

Authoring tool. Runs with you present, so asking questions here is fine — the goal is to produce a heavy prompt that is then launched-and-left with zero questions. Work in your project repo (current working directory). Paths below are relative to its root (the library lives under heavy/).

Step 1 — Load the conventions

Read these so the new prompt matches exactly:

• heavy/library/_TEMPLATE.md — the structure every prompt must follow (frontmatter + 7 sections).
• heavy/README.md — conventions (anchors, append-only, shared vocab, no-AskUserQuestion-mid-run, subagents off).
• heavy/library/INDEX.md — to find the next ID (highest existing NN + 1, zero-padded).

Step 2 — Interview the user

Use AskUserQuestion (or plain prose Q&A) to settle, in this order — batch related questions:

1. Goal — the one outcome the run must produce (if $ARGUMENTS was given, start from that).
2. Context — what problem it serves; which repo files/areas it touches.
3. Shape — roughly what phases/stages the work has (you'll refine into steps).
4. Subagents — almost always false (default; keeps the whole run in one session). Only true if the user explicitly wants subagents and accepts that spawned agents add complexity and may bill differently from the interactive session.
5. WebSearch — does it need the web? set a sane websearch_max (0 if not).
6. Cost — rough est_cost (low/medium/high) from the scope.

If the user is vague, propose a concrete decomposition and confirm it — don't leave it open.

Step 3 — Decompose into phases → steps

• Break the job into ordered phases; each phase = a list of idempotent steps.
• Give every step an id <phase>.<step> and a single defined output (a file + an anchored section <!-- SEC:id -->). Use the shared vocabulary where it fits: summary, findings.*, decisions, actions, assumptions, sources.
• The last phase is always synthesize (write/finish 00-summary.md, regenerate MAP.md, update heavy/output/INDEX.md).

Step 4 — Encode defaults for EVERY decision

This is the most important step. The run is unattended, so for any choice the prompt would otherwise ask about, write the default into the prompt plus the rule "when uncertain, take the default and record it under <!-- SEC:assumptions -->." A heavy prompt that needs a human mid-run is broken.

Step 5 — Write the prompt file

Copy _TEMPLATE.md structure into heavy/library/NN-slug.md (slug = short kebab-case of the goal), filled in: frontmatter (id, title, goal, output_dir: heavy/output/NN-slug/, phases, subagents, websearch_max, est_cost) + GOAL, CONTEXT, PHASES, OUTPUT SPEC (anchored, append-only), GUARDRAILS, RESUME (read STATUS.json), DONE (checkable criteria). Be fully self-contained — every path, every default, every output file named.

Step 6 — Register it

Add a row to heavy/library/INDEX.md (ID, Title, Goal, Cost, Subagents, Output dir).

Step 7 — Hand off to review

Tell the user the prompt is drafted and recommend /heavy-review NN to catch logical gaps before a real run. Offer to run the review now. Do NOT launch the heavy prompt itself from here — that's /heavy NN.

Do not commit unless the user asks (or follow the repo's normal commit discipline if they do).