plan-creation

Plan Creation Skill

Research the codebase, validate technical claims, and emit a work-ready spec.json in a single pass. Output validates against flywheel/schemas/task-list.schema.json.

Core Principles

1. Codebase reality first — dispatch locators before hypothesizing patterns.
2. Decisions, not code — capture approach, boundaries, risks, test scenarios. Do not pre-write implementation code.
3. Executable-from-day-one — if you can't commit to concrete file paths or test scenarios, surface an open_question instead of a vague task.
4. BLOCKING: Validate high-risk claims — security, payments, crypto, migrations, privacy trigger external validation via Context7.

Input

Feature description via $ARGUMENTS. If empty, ask user.

---

Phase 0: Check for Existing Knowledge

Before codebase research, check existing knowledge (skip missing dirs):

1. Standards (docs/standards/) — Search by tags for reusable patterns.
2. Solutions (docs/solutions/) — Verified fixes from past work (up to 5 matches).
3. Research (docs/research/) — Recent research within 30 days:

   find docs/research -name "*<topic-keywords>*" -mtime -30 2>/dev/null | head -3
   

If relevant knowledge found, use it as starting point for Phase 1. Fold key references into context.constraints[] or context.patterns[] so they survive into the dispatch.

---

Phase 1: Understand Codebase Context

BLOCKING: Do NOT use Read/Grep/Glob for TARGET CODEBASE research — dispatch locator Tasks first, then feed results to analyzer Tasks. Skill references, plan artifacts, and template files are exempt.

Run the canonical research workflow: all four locators in parallel → consolidate → all four analyzers in parallel. Read flywheel/skills/flywheel-conventions/references/research-workflow.md for locator templates, the consolidation rule, and analyzer templates.

After the canonical workflow completes, also check CLAUDE.md for team conventions (if analyzer-docs didn't already surface it) and recent similar features for precedent.