mcpField
D
diverse-plan

diverse-plan

github.com/robsonrungAgent workflows254 installsUpdated 2d ago
Repository

About diverse-plan

diverse-plan is an Agent skill for comparing materially different design premises and producing one executable implementation plan. It uses blind branches, targeted critiques, and a bounded completeness check, with work governed by an approved route snapshot.

Key features

  1. 01Compare distinct design premises
  2. 02Explore options through blind branches
  3. 03Run targeted critiques when needed
  4. 04Synthesize and check an implementation plan
  5. 05Record routes and execution seats

Use cases

  1. 01Implementation choices depend on alternatives
  2. 02Assess architecture or failure-handling options
  3. 03Create plans with files, tests, and risks

Install

Copy the command and run it in a terminal.

$npx skills add https://github.com/robsonrung/rar-skills --skill diverse-plan

After installing, type/diverse-planin Claude Code to confirm it loaded.

Bundled filesagents/openai.yaml

Details

From the project README

Diverse Plan

Choose the branch before selecting a model. Select branch premises and material review needs in the current coordinator. Before dispatching selected branches or reviewers, use shared/references/model-preview.md to bind one approved route snapshot.

Use this only when alternative designs can change the implementation decision. It is separate from models-consensus: the user chose a planning exercise, not a council. Never invoke the council from this skill.

Outcome

Return one executable plan. It names the chosen approach, rejected alternatives, files, tests, risks, and the effective seats used. A branch is useful only when it explores a materially different premise.

Routing

For selected workers, read task-shaped-model-routing.md, model-roster.md, and host-model-execution.md to map the snapshot roles. Use the selected semantic route, not a legacy seat name or pinned model id.

Map each selected branch to deep-analysis, architecture, or technical-analysis from shared/model-routing.json. Use precision-review only when a specific semantic risk needs an independent critique. Record the exact routes in the approved snapshot. Branches must explore different premises; they must not repeat the same brief at higher effort.

Use two blind branch seats by default. Add a third only when an unresolved boundary, failure case, or delivery tradeoff needs a distinct approach. The configuration supplies models and effort; this skill supplies the branch protocol.

Use the snapshot's exact native route through the active host first. Start each blind branch in its own isolated persistent subagent or supported task context. Use a runner only for a foreign model or a native route that cannot meet the approved plan, then capture its own session for later critique or synthesis turns. Probe only the selected capabilities. Preserve seat fidelity: a missing seat is unavailable, never silently replaced. Continue only with two distinct available branch seats. If that quorum is unavailable, return the missing prerequisites and the single-model to-prd path instead.

Route Approval

After branch selection, present the exact route table: branch or lens, seat, current model, effort, role, execution path, call count, and unavailable action. Include the blind branches, selected critiques, synthesis, and execution completeness check. Use this table for the shared preview and apply its decision rules once. Do not launch an unselected branch or critique. Record the approved snapshot with the plan artifacts and use only those routes.

Workflow

  1. Write one neutral brief with the goal, scope, constraints, acceptance contract, relevant files, and what a plan must decide. Ask one question only when different answers would produce different plans.
  2. Launch the blind branches with the same brief. Give each one premise: smallest viable change, cleanest design, or most robust failure handling. Each returns approach, files, tradeoff, risks, tests, and rollback point. Keep that branch context for its later critique and reply; a new opening context is a different branch.
  3. Compare branches against the acceptance contract. Run only the relevant engineering lenses. Each critique must identify the strongest point, material weakness, and a concrete test or check.
  4. Synthesize the record into one plan. Every adopted element traces to a branch or resolved critique. Preserve a material unresolved disagreement as a risk or open question.
  5. Run one bounded execution-completeness check. Patch the plan once if it exposes a material omission. Do not create an open-ended planning loop.

Synthesize and enrich

Use the synthesis route from the approved snapshot. It chooses the plan spine, incorporates verified improvements, and records why the strongest alternative lost. Keep the synthesis context separate from each blind opening. If the selected route is unavailable, stop and report the missing route or use only a user-approved alternate.

Execution completeness check

Use the implementation-review route from the approved snapshot to test whether an implementer can follow the plan without inventing a design decision. Check scope, file placement, acceptance criteria, commands, tests, rollout risk, and rollback. One material gap gets one targeted correction round; otherwise stop.

Output

Present:

  1. the change and chosen approach;
  2. approaches considered and the deciding tradeoff;
  3. ordered implementation steps, files, tests, and verification;
  4. risks, rollback, and open questions;
  5. the seat and lens table, including effective receipts, unavailable seats, and fallbacks.

Write an artifact only when the user asks. Hand the approved plan to to-prd so it becomes the workflow's approved product record; then to-tasks slices that PRD. Do not implement it here.

Comments

Markdown supported