Plan Feature

Plan Feature

Two-phase pre-implementation planning: clarify requirements via interview, then compile a complete PRD.

Phase 1: Discovery

Help clarify requirements and scope through structured questioning.

Interview Framework

The interview is a fog-of-war walk across four quadrants of the unknown. Open by listing the known knowns (what the user has already decided), then work the questions below to surface known unknowns (open questions they're aware of), unknown knowns (constraints they hold but haven't said — the questions in 2–4 exist to shake these loose), and close by hunting unknown unknowns ("what would surprise us mid-build? what reference implementation should we read first?"). A discovery that ends with all four quadrants visited produces a PRD that doesn't get re-planned in week two.

Goal Quality Bar (gate before interviewing)

Before opening the interview questions below, try to state the goal in one line that answers five things:

• What's true when this is done?
• What evidence shows it — a command, a test, a metric, a reviewed artifact?
• What threshold counts as success — pass/fail, or a number?
• What's explicitly out of scope, where that would matter?
• What's the stop condition — the point where you ask the user instead of guessing?

If a clean one-liner falls out, skip straight to Phase 2's clarifying questions — the interview below exists for when it doesn't.

Reject pure activity goals — "make progress," "keep investigating," "improve things" — until sharpened into a verifiable outcome:

• Weak: "Make checkout faster." Sharpened: "Reduce checkout API p95 latency below 250ms for the documented slow path, verified with npm run test:checkout and 3 consecutive benchmark runs under 250ms."
• Weak: "Clean up the auth code." Sharpened: "Resolve the open change-request threads on PR 123, touching only the affected auth files and tests, verified with the targeted auth test command plus gh pr view 123 showing no unresolved threads."

Pick the validator shape by domain:

Goal Quality Bar adapted from openai/skills define-goal, Apache-2.0.

1. Understand the Goal

• What problem are you solving?
• Who is this for?
• What does success look like?

2. Define Scope

• What must be included (MVP)?
• What's nice to have (future)?
• What's explicitly out of scope?

3. Identify Constraints

• Timeline constraints?
• Technical constraints?
• Resource constraints?

4. Clarify Details

• What are the inputs/outputs?
• What are the edge cases?
• What are the error scenarios?

5. Validate Understanding

• Summarize back what you heard
• Confirm priorities
• Identify open questions

Output

## Discovery Summary: [Feature/Project]

### Goal
[Clear statement of what we're building and why]

### Requirements
**Must Have (MVP)**
- [ ] Requirement 1
- [ ] Requirement 2

**Nice to Have**
- [ ] Feature A
- [ ] Feature B

**Out of Scope**
- Not doing X
- Not doing Y

### Technical Approach
[High-level approach]

### Open Questions
- [ ] Need to clarify: ...
- [ ] Decision needed: ...

### Next Steps
1. [First action]
2. [Second action]

Remember

• Ask, don't assume
• Summarize frequently
• Document decisions
• Store requirements as learnings

Phase 2: PRD Compilation

Structured 6-phase process to produce a complete PRD from a feature idea, including user stories, task breakdown, and parallel execution plan.

Workflow

Phase 1: Clarifying Questions

Ask 5-8 targeted questions to fill gaps. Use smart defaults so the user can skip.

## Clarifying Questions

1. **Target users?** [default: existing app users]
2. **Platform scope?** [default: web only]
3. **Auth required?** [default: yes, existing auth]
4. **Performance targets?** [default: <2.5s LCP, <200ms INP]
5. **Accessibility level?** [default: WCAG 2.1 AA]
6. **Data persistence?** [default: existing database]
7. **Mobile responsive?** [default: yes]
8. **Analytics needed?** [default: basic events]

Press enter to accept all defaults, or answer specific questions.

Phase 2: Scope Definition

Define what is IN and OUT of scope.

## Scope

### In Scope
- [feature 1]
- [feature 2]

### Out of Scope
- [explicitly excluded 1]
- [explicitly excluded 2]

### Assumptions
- [assumption 1]
- [assumption 2]

Phase 3: User Stories

Write user stories with acceptance criteria.

## User Stories

### US-1: [Title]
**As a** [role]
**I want** [capability]
**So that** [benefit]

**Acceptance Criteria:**
- [ ] Given [context], when [action], then [result]
- [ ] Given [context], when [action], then [result]

**Priority:** P1 | P2 | P3
**Complexity:** trivial | small | medium | large | epic

Phase 4: Task Breakdown

Break into implementable tasks with metadata. Dependencies/Blocks feed the batch algorithm and must match the Functional DAG's joins from Phase 5 — on conflict the DAG wins, so reconcile the metadata to it rather than redrawing to match a stale field.

## Tasks

### T-1: [Title]
- **Description:** [what to implement]
- **User Story:** US-1
- **Priority:** 1 (1=highest, 5=lowest)
- **Complexity:** medium
- **Estimated Tokens:** 15000
- **Dependencies:** none
- **Blocks:** T-2, T-3
- **Verification:** [how to verify completion]

### T-2: [Title]
- **Dependencies:** T-1
- ...

Phase 5: Parallel Batch Detection

Draw the Functional DAG first (docs/functional-dag.md) — inputs left, operations merging rightward, one terminal verification node — then read the batches off its columns. Same column with disjoint inputs = same batch. The DAG is the source; the batch list below is its rendering, not a second dependency graph to keep in sync.

## Functional DAG

[recipe-table brace diagram in a fenced code block — see docs/functional-dag.md]

## Execution Plan

### Batch 1 (parallel) — estimated: 25k tokens
- T-1: Setup data models
- T-4: Create UI scaffolding
- T-7: Write test fixtures

### Batch 2 (parallel, depends on Batch 1) — estimated: 40k tokens
- T-2: Implement API endpoints (depends: T-1)
- T-5: Build form components (depends: T-4)

### Batch 3 (sequential) — estimated: 20k tokens
- T-3: Integration wiring (depends: T-2, T-5)

### Batch 4 (parallel) — estimated: 15k tokens
- T-6: E2E tests (depends: T-3)
- T-8: Documentation (depends: T-3)

### Batch 5 (terminal gate) — estimated: 5k tokens
- T-9: typecheck + full test run (depends: T-6, T-8) — the DAG's single terminal node

Total estimated tokens: 105k
Estimated context windows: 2

Every plan ends on the terminal gate, so the last batch is always one verification task depending on all preceding work — never a fan-out of unverified parallel tasks.

Phase 6: Final PRD

Compile everything into the final document.

# PRD: [Feature Name]

## Functional DAG
[from Phase 5 — inputs left, operations merging rightward, one terminal node]

## Overview
[1-2 paragraph summary]

## Goals
- [measurable goal 1]
- [measurable goal 2]

## Scope
[from Phase 2]

## User Stories
[from Phase 3]

## Technical Design
### Architecture
[high-level approach]

### Data Model
[key entities and relationships]

### API Surface
[endpoints or interfaces]

## Task Breakdown
[from Phase 4]

## Execution Plan
[from Phase 5]

## Success Metrics
- [metric 1: target value]
- [metric 2: target value]

## Risks
| Risk | Likelihood | Impact | Mitigation |
|------|-----------|--------|------------|
| [risk] | Low/Med/High | Low/Med/High | [approach] |

## Timeline
- Phase 1: [dates] — [deliverable]
- Phase 2: [dates] — [deliverable]

Smart Defaults

When the user provides minimal input, apply these defaults:

• Platform: Web — detect from package.json (next → Next.js / satus, react-router → RR / novus)
• Auth: Existing auth system
• Performance: LCP <2.5s, INP <200ms, CLS <0.1
• Accessibility: WCAG 2.1 AA
• Testing: Unit + integration, no E2E unless requested
• Styling: Project's existing system (Tailwind if detected)

Task Sizing Reference

See docs/enhanced-todos.md for the complexity/token sizing reference table.

Phase 3: Close (plan → durable record)

A PRD is a build plan while building and a rationale record once shipped — closing flips it from one to the other. When the feature lands (merged, gates green), rewrite the PRD instead of letting it rot as a stale plan:

• Keep the why: the problem, the principles, the invariants that must never break, the approaches tried and rejected.
• Cut every paragraph that restates what the code does — point at the code instead; it is the source of truth for how.
• Record every divergence from the plan (dropped tasks, renamed seams, assumptions that broke). Divergences are the most valuable content: they're exactly what a future reader would otherwise re-derive the hard way.
• Archive it where the project keeps finished plans (e.g. docs/prd/done/), and promote anything decision-shaped into an ADR via /context-doc.

If the plan lives as GitHub issues (/project), the close pass is the closing comment on the epic: what shipped, what diverged, and why.