Graph Plan

---
name: graph-plan
description: Design and render the DAG for a goal — nodes, edges, schemas, cost
  estimate, and the full workflow script — WITHOUT executing anything. The dry-run
  companion to /graph.
argument-hint: <goal>
disable-model-invocation: true
---

# Graph-plan: design the DAG, spend nothing

Design the graph for this goal, but **never execute it**:

> $ARGUMENTS

Do NOT call the Workflow tool. Do NOT call the Agent tool. This is a pure design
pass — its entire value is letting the user see the graph before spending tokens.

## Procedure

Follow Steps 0–3 of the `graph` skill (read `../graph/SKILL.md` if not already in
context, and `../graph/reference/patterns.md` for the shapes):

- **Step 0 — Qualify.** If the goal shouldn't be a graph, say so and what you'd
  do instead. That's a complete, useful answer; stop there.
- **Step 1 — Decompose** into verb-first nodes and layers.
- **Step 2 — Contract** the edges with JSON Schemas and write the node briefs.
- **Step 3 — Choose the shape(s)** from the pattern gallery.

Then, before writing the output script, apply the Step 4 Path A authoring
checklist from `../graph/SKILL.md` and read `../graph/reference/workflow-api.md`
— but do NOT invoke the tool. The script you print must be the real, runnable
artifact, not pseudocode.

## Output

Present, in order:

1. **ASCII DAG** — layers, arrows, parallel groups marked.
2. **Node table** — `node | job (one line) | effort | feeds into`.
3. **Edge schemas** — each schema, compact.
4. **Cost estimate** — projected agent invocations, computed the Step 0 way:
   sources + Σ(schema `maxItems` × lenses × votes) + loop worst case. Give the
   band (small <15 / medium 15–40 / large >40 agents) and flag the ~25-agent
   native warning threshold if crossed.
5. **The full workflow script** in a fenced `js` block — exactly what `/graph`
   would run, ready to eyeball or hand-edit.

Close with: *"Run `/graph <same goal>` to execute this, or `/graph-save <name>`
after a run to keep it as a `/command`."*