plan-feature（计划特征）

Feature Planning

Create concrete, implementation-ready plans for features and complex changes.

DO NOT WRITE CODE during planning. Only explore, analyse, and document.

Planning threshold

Before deep planning, quickly confirm the scope:

• Plan in detail when changes affect 3+ files or multiple systems
• Plan in detail when requirements are unclear or architectural tradeoffs exist
• Skip heavyweight planning for single-file changes with clear requirements
• For small/obvious changes, give a short execution outline instead of a full spec

Planning workflow

1. Discovery

Ask targeted questions to uncover intent. For each question:

• Present 2-3 concrete options with tradeoffs
• Give your recommended option with clear reasoning
• One question at a time; wait for user response
• Skip questions already answered by the user

Critical questions:

• What problem are you solving? (user pain point, business goal)
• What should happen? (expected behaviour, success criteria)
• What should NOT happen? (constraints, edge cases to avoid)
• Who is this for? (user type, environment, scale)
• How will you verify it works? (testing approach, validation)

Speed-to-learning reference

• Use ship-fast-loop.md for a lightweight shipping loop and feedback cadence.

2. Analysis

Explore the codebase systematically:

• Locate relevant files (document paths with line numbers)
• Map existing patterns (architecture, naming, data flow)
• Identify dependencies (what will be affected by changes)
• Find similar implementations (to maintain consistency)
• Note relevant standards (from implement-frontend, define-architecture, etc)

Document findings:

• File: path/to/file.ts:123 - what it does, how it's relevant
• Pattern: existing approach for similar features
• Constraint: technical limitation or requirement

3. Planning

Create a concrete, ordered plan with:

For each change, specify:

• File path and approximate line number
• Exact function/component/class to modify
• What to add/remove/change (be specific)
• Why this change (how it fits the goal)
• Dependencies (what must happen first)

Plan structure:

## Goal
[One sentence: what we're building and why]

## Changes

### 1. [Description]
- File: `path/to/file.ts:45`
- Action: Add `functionName()` that does X
- Reasoning: Needed because Y
- Dependencies: None

### 2. [Description]
- File: `path/to/other.ts:89`
- Action: Modify `existingFunction()` to handle Z
- Reasoning: Integrates with change #1
- Dependencies: #1 must complete first

## Validation
- [ ] Tests pass
- [ ] Feature works for case A
- [ ] Edge case B is handled
- [ ] Follows `implement-frontend` (if frontend)
- [ ] No console logs or debug code

4. Standards reference

Explicitly note which standards apply:

• Frontend changes: reference implement-frontend, audit-ui
• UI changes: reference design-ui
• Motion: reference ui-animation
• Backend: reference define-architecture
• Typography: reference audit-ui

Format: "This plan must follow implement-frontend for forms and type safety."

Anti-patterns

Avoid vague plans:

• Bad: "Update the authentication system"
• Good: "Modify auth/middleware.ts:34 to add validateSession() that checks token expiry"

Avoid missing context:

• Bad: "Add error handling"
• Good: "Wrap API call in auth/api.ts:67 with try/catch, show toast on error per audit-ui"

Avoid assuming knowledge:

• Bad: "Use the standard pattern"
• Good: "Follow the existing DAO pattern from user/dao.ts:12 (class-based with explicit types)"

Avoid incomplete acceptance criteria:

• Bad: "Make sure it works"
• Good: "Verify: (1) form submits on Enter, (2) shows inline errors, (3) disables submit during request"

Avoid ignoring standards:

• Bad: Plan uses any types and manual form state
• Good: Plan enforces implement-frontend: no any, uses React Hook Form

Validation checklist

Before handing off the plan, verify:

• Someone can implement without asking questions
• File paths are valid with approximate line numbers
• Steps are ordered with clear dependencies
• Acceptance criteria are testable
• Relevant skill standards explicitly referenced
• Every decision is justified (the "why")
• No ambiguous language ("update", "improve", "enhance" without specifics)