Create Plan

Create Plan

Hard rules

• Plan only — never implement.
• Allowed writes: .plans/<kebab-name>.md, and .reviews/*-decomplex.md when needed.
• Evidence-backed. Smallest design. No speculative scope.
• Ask user on material ambiguity (scope/behavior/architecture/migration/risk/complexity). No shaky assumptions.
• Delegate by default for research and review when safe. “Small/easy” ≠ skip subagents.
• Leverage subagents — built-in, extensions/plugins, or skills. Follow use-subagents policy; use native subagents, extensions / plugins, or other skills
• Main agent owns synthesis, plan writes, dispositions, delivery.
• Child/reviewer output = evidence, never acceptance. Findings never auto-enter the plan.
• Prefer simpler fixes. Stuck review (2 failed rounds / recurrence / no progress) → ask user.

Load or stop

• Before draft: read assets/implementation-plan-template.md.
• Before finalize (and after material edits): read references/plan-quality-checklist.md.
• Missing resources → stop. Don’t invent substitutes.

Tasks must have

• Stable IDs
• Change as bullets (one concrete edit/behavior each)
• Starts-at paths/symbols (non-exhaustive)
• Exact verify commands + expected signals
• Each behavior-changing task names an existing/planned test path or suite, the observable behavior, boundary, or regression it protects, and the narrowest credible repository-conventional layer; or justifies omitting automation and gives alternate validation
• Omit empty optionals
• Key files table: path · why · plan impact

Sequence

1. Frame — request, repo instructions, existing plans → outcome, scope, non-goals, risks, validation. Don’t deep-explore in parent when a research lane can.
2. Load template.
3. Research — enumerate questions; delegate by default (one question/scope/stop per lane). Inspect relevant existing tests, framework/configuration, and repository commands. External only if decision-relevant. Verify critical claims.
4. Resolve gates — still ambiguous? ask user before drafting as fact.
5. Synthesize — smallest approach; map findings → decisions/tasks/checks/non-goals/gates; drop fluff.
6. Draft — adapt template; flat tasks default; phases only for real boundaries; no review ledgers in the handoff.
7. Complexity — structural draft → decomplex Prevention if available, else built-in gate. Complexity-increasing accept → triage; doubt → ask user.
8. Review — consequential plans: delegate fresh independent reviewers → disposition (Accept/Validate/Reject/Ask user/Block) → revise → re-review until Clear. Small/unavailable: parent checklist + independence limit.
9. Checklist → deliver — save under .plans/ (or chat + no-write reason). Report path, gates, decisions, fallbacks, risks.

Stop

• Missing template/checklist → stop
• Required research/review unsafe/unavailable → block/escalate
• Persistent disagreement or uncertainty → ask user