PRD to Plan

PRD to Plan

Break a PRD into phased vertical slices (tracer bullets).

Interactive prompts: present options as a numbered list and wait for the user's choice.

Output: .tracerkit/plans/<slug>.md.

Input

The argument (if provided) is: $ARGUMENTS

Use argument as <slug>. If empty, list PRDs as numbered options and wait for the user's choice.

Workflow

1. Read the PRD

Read .tracerkit/prds/<slug>.md. If missing, list PRDs as numbered options and wait for the user's choice.

If .tracerkit/plans/<slug>.md exists, present options and wait:

1. Overwrite existing (Recommended)
2. Pick a new name

2. Explore the codebase

Map architecture, patterns, integration points. Skip if codebase context exists from prior step.

Research protocol: codebase first, then docs. Unverifiable claims → flag as uncertain, never fabricate.

3. Identify durable architectural decisions

Before slicing, extract decisions that hold across all phases:

• Route structures / URL patterns
• Database schema shape
• Key data models and definitions
• Auth/authorization approach
• Third-party service boundaries

4. Draft vertical slices

Each phase: thin vertical slice through all layers (schema → service → API → UI → tests). Demoable alone.

Deriving tasks from the PRD:

PRD Section	Becomes
New Modules	Implement module with interface
Schema Changes	Migration + validation
API Contracts	Route returning shape
Navigation	Wire component to route
User Stories	Verify coverage; add task if missing
Testing Decisions	Tests land in the phase where their module lands
Out of Scope	Never create tasks for these

Within each slice, order by dependency: schema → service → API → UI → tests. Happy paths before edge cases.

Phase naming: use a goal phrase answering "what can we demo when this is done?" (e.g., "Phase 1 — Revenue visible end-to-end"), not a layer name.

Done when: checkbox list of atomic, verifiable conditions. Each must name a test file/name, a shell command, or a file+content to verify. No prose-only conditions. Test: "Can an agent verify by reading files, running a command, or checking a test?"

Layer-by-layer exception: if complex schema changes underpin all modules and no story stands alone, build data foundation first, then slice vertically.

Phase count thresholds:

• 1 module touched → 2–3 phases max
• 2–3 modules touched → 3–5 phases max
• 4+ modules or 6+ phases → stop and present options:
  1. Split the PRD (Recommended)
  2. Continue anyway

Count "modules touched" by scanning the PRD's New Modules and Schema Changes sections.

Assign an agent tag to tasks where appropriate:

• [agent:debugger] — tracing a bug or unexpected runtime behavior
• [agent:test-auditor] — writing or reviewing tests
• [agent:code-reviewer] — reviewing API surfaces, interfaces, or public contracts

5. Quiz the user

Present breakdown (title, user stories covered, done-when per phase). Present options and wait:

1. Looks good, proceed (Recommended)
2. Merge some phases
3. Split a phase

Iterate until approved.

6. Save plan

Save to .tracerkit/plans/<slug>.md (create dir if missing).

# Plan: <Feature Name>

Use this structure for the plan body:

## Architectural Decisions

Durable decisions that apply across all phases:

- **Key decision**: ...

---

## Phase 1 — <Goal>

**User stories**: <list from PRD>

### What to build

Concise description of this vertical slice — end-to-end behavior, not layer-by-layer.

### Done when

- [ ] Atomic, testable condition
- [ ] Another testable condition

---

<!-- Repeat for each phase -->

## Out of Scope

Carried forward from PRD verbatim.

## Open Questions

Gaps found in the PRD needing resolution. Blank if none.

Print one line per phase: Phase N — <title> (<condition summary>). Present options and wait:

1. Run /tk:build <slug> (Recommended)
2. Run /tk:check <slug>
3. Done for now

Execution guidance

To implement this plan phase by phase, run /tk:build <slug>. It handles branch creation, implementation, feedback loops, checkbox marking, and commits — one phase per invocation.

Rules

• Phases derive from PRD user stories — never invented
• Each phase must be demoable end-to-end on its own
• "Done when" must be a checkbox list of testable conditions, not prose
• Safety valve: if a phase has >5 "Done when" items, stop and split it into smaller phases before continuing
• Never modify the source PRD content
• Carry PRD's Out of Scope forward verbatim