diverse-plan
Installation
SKILL.md
Diverse Plan
Use this only when alternative designs can change the implementation decision. It is a planning exercise, not a council; never invoke models-consensus from here. The result is one executable plan.
Plan
- Write one neutral brief: goal, scope, constraints, acceptance contract, relevant files, and what a plan must decide. Ask one question only when different answers would produce different plans.
- Choose the branches. Two by default; add a third only when an unresolved boundary, failure case, or delivery tradeoff needs its own approach. Each branch gets one premise: smallest viable change, cleanest design, or most robust failure handling. A branch that repeats another's brief at higher effort is not a branch.
- Resolve each branch with
python3 shared/scripts/model_routing.py resolve <route> --provider <host>, usinganalysisordeep-analysis(defined inshared/model-routing.json; shapes inshared/references/task-shaped-model-routing.md), andreviewonly for a named semantic risk. Branches run on distinct models. - Show the branches with their premise, model, effort, and host route, plus the synthesis and completeness-check roles. Ask: "Approve this plan, or specify changes."
Route each role by host: T3 Code delegate_task, continued with t3_thread_send (a Muse route uses muse-runner); Claude Code or Codex, a native subagent; otherwise cli-runner, pi-runner, or muse-runner. Each branch keeps one context for its later critique and reply; synthesis gets its own. Seat fidelity: an unavailable branch is reported, never replaced. With fewer than two distinct branches, stop and offer the single-model to-prd path.
Run
- Branches, blind and in parallel, from the same brief. Each returns approach, files, tradeoff, risks, tests, and rollback point.
- Critique against the acceptance contract, using only the lenses that apply: the strongest point, the material weakness, and a concrete check for each branch.
- Synthesis. Choose the plan spine and fold in verified improvements. Every adopted element traces to a branch or a resolved critique; record why the strongest alternative lost, and keep a material unresolved disagreement as a risk or open question.
- Completeness check, once: could an implementer follow the plan without inventing a design decision? Check scope, file placement, acceptance criteria, commands, tests, rollout risk, and rollback. Patch one material gap, then stop.