Define how the spec gets built, in small steps an agent can execute and verify one at a time. A good plan makes the build boring: every task is obvious, small, and checkable.
Input path: molten-docs/spec/<slug>.md from mt-spec. If missing or still draft, recommend finishing mt-spec first; proceed only if the user insists, and note the risk in the plan.
Output path (canonical): molten-docs/plan/<slug>.md. Create the directory if it does not exist.
Next skill: mt-build.
Operating Rules
- At the start of interactive runs, advise the user: "The best way to interact with this skill is to use voice mode to dictate your feedback."
- Read the codebase before planning. Name real files, modules, and patterns — not hypothetical ones.
- Plan in read-only mode: no edits to application code while planning.
- Slice vertically, not horizontally. Each task delivers one thin working path through the stack (UI → logic → data), not "all the database, then all the API, then all the UI".
- Size tasks to fit one focused agent session, touch roughly 1–5 files, and end in a verifiable state (tests pass, page renders, endpoint responds). If a task needs "and" to describe it, split it. If you cannot state its done-condition in one sentence, it is too big.
- Every task maps back to one or more spec acceptance criteria (AC-n). Every AC is covered by at least one task.
- Map dependencies, then order so foundations and the riskiest work come first and the app stays runnable between tasks.
- Put a checkpoint after each phase: full test run, manual smoke check, and a chance for the human to redirect.
- Include a test or verification step for each task.
- Ask every question as plain numbered text in chat; for finite choices, list lettered options under the question (1a, 1b, 1c). Do not use agent question tools (
AskUserQuestion,AskQuestion,request_user_input). - Do not start implementing while this skill is active — hand off once the plan is approved.
Workflow
- Read the spec, prototype learnings, and the relevant code.
- Identify architecture decisions and dependencies (new packages, migrations, env vars).
- Draft the dependency order, then group work into phases and break phases into checklist tasks.
- Review with the user — sequencing, scope per task, anything missing, anything that can be cut.
- Write the plan and point the user to the file. Set
status: approvedonly on explicit sign-off.
Output template
---
feature: <name>
slug: <slug>
created: <YYYY-MM-DD>
spec: molten-docs/spec/<slug>.md
status: draft | approved | in-progress | done
molten_next_skill: mt-build
---
# Plan — <name>
## Approach
## Architecture decisions
## Dependencies & setup (packages, migrations, env vars)
## Phase 1 — <name>
- [ ] **1.1 <task>** — files: `…` · covers: AC-1 · depends on: — · verify: <test/command/check>
- [ ] **1.2 <task>** — …
- **Checkpoint 1:** tests green, app runs, <manual check>
## Phase 2 — <name>
- [ ] …
- **Checkpoint 2:** …
## Risks & mitigations
## Definition of done (all acceptance criteria verified)
Done checklist
- Every spec acceptance criterion is covered by at least one task
- Each task names real files, a done-condition, and a verification step
- Tasks are vertical slices ordered by dependency so the app stays runnable
- A checkpoint closes each phase
-
molten-docs/plan/<slug>.mdwritten; status set toapprovedonce the user signs off
Extending
Common additions: effort estimates per task, parallelisable-task markers for multi-agent builds, rollout/feature-flag steps, or exporting tasks to Linear or GitHub issues.