| name | ultraplan |
| description | Generator-critic planning. Launches 3 parallel Opus 4.7 explorers with distinct divergence vectors (architectural, speed-first, simple-first), each producing an independent plan, then runs a critic synthesis that composes the best elements into a single plan with per-decision attribution. Defeats single-agent anchoring. Use when the user asks to plan a non-trivial task, design an approach, compare implementation strategies, or wants multiple angles on a problem before coding.
|
| arguments | [{"name":"problem","description":"The problem to plan. One or two sentences is enough.","required":false}] |
| user_facing | true |
Contents
| Reference | When to load |
|---|
references/orchestrator.md | Running /ultraplan <problem> — the 5-phase flow |
references/divergence-vectors.md | The three explorer biases and their motivation |
references/write-protocol.md | Copy verbatim into every explorer prompt |
templates/explorer-skeleton.md | Per-explorer plan skeleton |
templates/plan-skeleton.md | Final composed plan skeleton |
Ultraplan
Generator-critic planning, file-coordinated. Three Opus 4.7 explorers plan the same problem in parallel, each biased toward a different axis — architectural, speed-first, simple-first. A critic reads all three and composes a final plan that cherry-picks the best decision from each, with attribution. The user gets one clean plan.md to execute against.
Beats single-agent planning on two axes:
- No anchoring. A single agent commits to the first approach it reaches and defends it through the rest of the plan. Three independent explorers reach genuinely different conclusions, and the critic picks per-decision.
- Calibrated confidence. When all three converge, the approach is sound. When they diverge, the critic names the tradeoff explicitly.
How it works
Read references/orchestrator.md and follow its phases. Short version:
- Intake — parse the problem. One targeted
AskUserQuestion only if scope is genuinely ambiguous; otherwise proceed.
- Skeletons — create
planning/{slug}/ with three explorer files and a plan.md shell.
- Launch — three
general-purpose Opus 4.7 Agents in parallel, run_in_background: true, each with a distinct divergence vector from references/divergence-vectors.md. Orchestrator does not read the codebase; explorers do.
- Monitor — escalating check-ins (30s → 2m → 5m).
wc -l on explorer files; stuck detection via line-count delta; launch a fresh Agent with "resume from section N" context on stall.
- Critic synthesis — one foreground
general-purpose Opus 4.7 Agent reads all three explorer files and writes plan.md — chosen approach, per-decision attribution, explicit tradeoffs, step-by-step order, verification criteria.
Write discipline
Every explorer prompt carries the write protocol verbatim from references/write-protocol.md: think → edit the file with the decision → next consideration. Partial plans written to disk survive timeouts; plans held in working memory do not.
Cost note
Three parallel Opus 4.7 explorers plus one critic synthesis. Significant token spend per run — worth it for meaningful planning decisions, not for trivial refactors. The user invoked /ultraplan because they want the divergence; ship it.