architect（架构师）

Architect Skill - Implementation Plan Design

This skill provides architectural guidance for creating high-quality implementation plans. It focuses on universal principles applicable to any technology stack.

When to Use

• During the PLAN phase of AEP workflow
• When reviewing an existing implementation plan
• When designing a new feature architecture
• When refactoring or restructuring code

Core Principles

1. Atomic Phases

Each phase MUST be:

• Self-contained: Can be understood without reading other phases
• Independently deployable: Could theoretically be deployed alone
• Reversible: Can be rolled back without affecting other phases
• Testable: Has clear validation criteria

Rule of thumb: If a phase touches more than 3 files, it's probably too big.

2. Dependency Management

GOOD: Linear dependency chain
Phase 1 → Phase 2 → Phase 3

GOOD: Independent phases (can run in parallel)
Phase 1 ──┐
Phase 2 ──┼→ Phase 4
Phase 3 ──┘

BAD: Circular dependencies
Phase 1 ←→ Phase 2

Always ask: "Can this phase be completed if the next phase fails?"

3. Risk Ordering

Order phases by risk level (lowest risk first):

1. Infrastructure - New files, directories, configs (low risk)
2. Data structures - Types, interfaces, schemas (low-medium risk)
3. Core logic - Business rules, algorithms (medium risk)
4. Integration - Connecting components (medium-high risk)
5. Migration - Data changes, breaking changes (high risk)

4. Validation at Every Step

Each phase MUST specify:

• What to validate (tests, linter, manual check)
• Expected outcome
• How to verify success

Phase Design Checklist

Before finalizing each phase, verify:

[ ] SCOPE
    [ ] Does this phase have a single, clear goal?
    [ ] Are all changes directly related to that goal?
    [ ] Would splitting this phase improve clarity?

[ ] FILES
    [ ] Are there 3 or fewer files modified?
    [ ] If more, is there a strong reason?
    [ ] Are all file paths explicit and correct?

[ ] DEPENDENCIES
    [ ] Are all dependencies on previous phases explicit?
    [ ] Can this phase run if the next phase fails?
    [ ] Are external dependencies documented?

[ ] VALIDATION
    [ ] Is there a concrete way to validate this phase?
    [ ] Are tests specified (new or existing)?
    [ ] Is the expected outcome clear?

[ ] ROLLBACK
    [ ] Can changes be reverted without side effects?
    [ ] Are database migrations reversible?
    [ ] Is there a rollback procedure documented?

[ ] COMMIT MESSAGE
    [ ] Does it follow conventional commits?
    [ ] Does it accurately describe the change?
    [ ] Is it in English and imperative mood?

Plan-Level Checklist

Before finalizing the complete plan:

[ ] COMPLETENESS
    [ ] Does the plan fully address the original request?
    [ ] Are edge cases considered?
    [ ] Is error handling included?

[ ] ORDERING
    [ ] Are phases in optimal order (lowest risk first)?
    [ ] Are dependencies respected?
    [ ] Could any phases be parallelized?

[ ] TESTING STRATEGY
    [ ] Are unit tests included where appropriate?
    [ ] Are integration tests planned?
    [ ] Is manual verification documented?

[ ] DOCUMENTATION
    [ ] Will code changes require doc updates?
    [ ] Are API changes documented?
    [ ] Are breaking changes clearly marked?

[ ] RISKS
    [ ] Are potential risks identified?
    [ ] Are mitigations proposed?
    [ ] Is there a fallback if something goes wrong?

Universal Patterns

Pattern: Feature Flag Approach

For risky features, consider:

Phase 1: Add feature behind flag (disabled)
Phase 2: Implement feature logic
Phase 3: Add tests
Phase 4: Enable flag (or remove)

Benefit: Can ship incrementally, easy rollback.

Pattern: Strangler Fig

For replacing existing code:

Phase 1: Create new implementation alongside old
Phase 2: Route new requests to new implementation
Phase 3: Migrate existing usages
Phase 4: Remove old implementation

Benefit: Never breaks existing functionality.

Pattern: Database Migration Safety

For schema changes:

Phase 1: Add new column (nullable)
Phase 2: Dual-write to old and new
Phase 3: Backfill existing data
Phase 4: Make new column required
Phase 5: Remove old column

Benefit: Zero-downtime migrations.

Pattern: API Evolution

For API changes:

Phase 1: Add new endpoint/field (don't remove old)
Phase 2: Update consumers to use new
Phase 3: Deprecate old endpoint/field
Phase 4: Remove old (after deprecation period)

Benefit: Backwards compatibility maintained.

Pattern: Test-First Integration

For complex integrations:

Phase 1: Write integration tests (failing)
Phase 2: Implement minimal passing version
Phase 3: Refine and optimize
Phase 4: Add edge case handling

Benefit: Clear success criteria from start.

Anti-Patterns to Avoid

Anti-Pattern: Big Bang Phase

Bad:

## Phase 1: Implement the feature
- Modify 15 files
- Add database tables
- Create API endpoints
- Add frontend components

Fix: Split into focused phases by layer/concern.

Anti-Pattern: Hidden Dependencies

Bad:

## Phase 2: Add validation
(secretly depends on Phase 4's utility functions)

Fix: Make dependencies explicit or reorder phases.

Anti-Pattern: Optimistic Validation

Bad:

**Validation**: Should work

Fix: Specify concrete command: make test, npm run lint, etc.

Anti-Pattern: Implicit Rollback

Bad:

## Phase 3: Migrate user data
(no mention of what happens if it fails)

Fix: Add rollback procedure or mark as high-risk.

Anti-Pattern: Kitchen Sink Commit

Bad:

**Commit message**: `feat: add feature and fix bugs and refactor`

Fix: One concern per phase = one clear commit message.

Plan File Format

Plans MUST follow this format:

---
feature: {ticket-id-slug}
ticket_id: {ticket-id}
created: {ISO timestamp}
status: pending
total_phases: {N}
---

# {ticket-id}: Implementation Plan - {title}

## Overview

{Brief description of what will be implemented}

---

## Phase 1: {Phase Title}

**Goal**: {What this phase accomplishes}

### Files to modify/create

- `path/to/file.ext` - Description of changes

### Validation

{Concrete command: make test, phpunit tests/Specific/, npm run lint, etc.}

### Commit message

{conventional commit: feat|fix|refactor(scope): description}

---

## Phase 2: {Phase Title}

...

---

## Summary

| Phase | Description | Files | Complexity |
|-------|-------------|-------|------------|
| 1 | ... | N | Low/Medium/High |
| 2 | ... | N | Low/Medium/High |

Important:

• Frontmatter is REQUIRED for automation
• Phase headers MUST follow format: ## Phase N: Title
• Commit messages MUST be in code blocks
• Validation MUST be concrete commands, not "should work"

Output Format

When using this skill during planning, structure architectural decisions as:

## Architectural Decisions

### Decision 1: [Title]
- **Context**: Why this decision is needed
- **Options considered**: What alternatives exist
- **Decision**: What was chosen
- **Rationale**: Why this option
- **Consequences**: What this implies for implementation

### Phase Structure Rationale
- Why phases are ordered this way
- Which phases could potentially be parallelized
- Where the highest risk lies

Integration with AEP

This skill is designed to be invoked during AEP's PLAN phase:

1. AEP completes ANALYSE and EXPLORE
2. Before creating the plan, invoke Architect skill
3. Apply checklists to each phase
4. Validate plan structure against anti-patterns
5. Document architectural decisions
6. Output final plan with confidence

Language

ALL OUTPUT MUST BE IN FRENCH when communicating with the user. Technical terms and commit messages remain in English.