/create-plans — PRD → execution plans

You are a thin orchestrator. Invoke each skill by name with the stated focus; one-line progress summary at each step boundary.

Two standing rules, because sub-skill output buries the parent sequence and compaction silently drops instructions:

• After EVERY skill returns, restate position: "→ Returning to /create-plans (unit X): next is step N (<name>)."
• After any auto-compaction, re-invoke the skill governing the current step before continuing.

Input: $ARGUMENTS — PRD name, Initiative name, or vault path. Resolve it under C:\Users\any0ne\vault\openlatch-vault\03-Product\PRDs\. Build the unit list: the PRD itself if it has no Initiatives, otherwise every Initiative — ordered by their depends-on DAG. A single Initiative as input is valid: the unit list is just that Initiative — but check its depends-on first and warn Luc if an upstream unit has no plans yet (its plans would build on assumptions that don't exist).

Goal: set a session goal now (SlashCommand /goal if available, otherwise state it and treat it as binding): "Every unit has a brainstorm folder (design.md + 1–4 plan files + INDEX + outbox) in its target repo's `.local/.brainstorms`, adversarially reviewed and verified one-shot-ready; Index.md status → Planned."

Per unit (in DAG order)

1. Skill: `calibrate` — focus plans. Returns the plan-file count (1–4; fewer is better) and the Tier; derive step 4's question budget from the Tier (Light ≥10 / Medium ~20 / Heavy 35+).
2. Skill: `learn-existing` — focus prd:<unit>. Direct read of the PRD + this Initiative (primary artifacts — verbatim fidelity matters).
3. Skill: `custom-deep-research` — focus technical: 2026 best practices, designs, and frameworks for this unit's domain.
4. Skill: `learn-existing` — focus codebase of the unit's target-repos, informed by the research (where this lands, conventions, reusable pieces, constraints).
5. Skill: `clarify-intent` — focus: technical design decisions (architecture choices, trade-offs, integration points). Budget from step 0's Tier.
6. Author the brainstorm folder at <repo>/.local/.brainstorms/<kebab-slug>/:

• YYYY-MM-DD-design.md — context, source-of-truth pointers table, prime invariant, decisions, risks (Risk | Mitigation table), stack — synthesized from steps 1–4.
• plan/INDEX.md — decision registry (D-XX entries) + per-plan status.
• plan/01..0N-<phase>.md — each one-shot executable: decision-ownership header (> Owns: D-XX…), Scope, per-module sections with explicit algorithms as commented code blocks, REJECTED-alternatives gotchas, Tests, Files table (File | Treatment | Notes), Acceptance (observable outcomes, curl-able where possible), Rollback.
• outbox/ — walkthroughs for actions only Luc can do (account setups, registrations).
• Record the folder path in the unit's brainstorm-folders frontmatter (the Initiative's, or the PRD's own when it has no Initiatives). Create 03-Product/PRDs/Index.md if missing, then set this PRD's row to Planned.

1. Skill: `adversarial-review` — focus: the plan files just written. Integrate genuine gaps; discard over-engineering.
2. Verification workflow — author a dynamic Workflow: one fresh-context agent per plan file answering "could I execute this with NO other context?" (file paths real against the repo? commands accurate? steps ordered? gotchas explicit? acceptance checkable?) plus one cross-plan agent (plans compose with no overlap/gap vs the PRD). Fix findings, re-verify changed files.

Done

Per-unit summary line as you go; finish with: folders created (paths), plan counts and why, open outbox items for Luc, and the next command: /execute-plans "<PRD name>".