Plan

Implementation Planning

Quick Reference

Approach	When to Use	Artifact
Direct execution	Small change, <3 files	Code
Lightweight plan	Medium feature, 3-5 files	Mental model + steps
Full written plan	Large feature, 5+ files	Plan document
Interview-then-execute	Unclear requirements	Q&A session, then plan
Prototype first	Exploratory / uncertain	Working prototype

Decision Tree

How big is this change?
├── 1-2 files, well-defined → Just do it (no plan needed)
├── 3-5 files, clear scope → Lightweight plan in conversation
├── 5+ files or architectural → Full written plan
├── Requirements unclear → Interview first
└── Feasibility unknown → Prototype first

When to Plan Formally

• Architectural changes (new auth system, database migration)
• Multi-service changes (frontend + backend + database)
• Changes where ordering matters (migrations before code)
• When you need team alignment before coding
• Unfamiliar codebases -- planning forces understanding first

What a Plan Should Contain

1. Scope -- what's changing and what's not
2. Steps -- numbered, ordered by dependency
3. Files -- specific files to create or modify per step
4. Verification -- how to test each step
5. Risks -- what could go wrong, rollback strategy

Plan Template

## Plan: [Feature Name]

### Scope
- Adding: [new capabilities]
- Modifying: [existing files/behavior]
- Not touching: [explicitly out of scope]

### Steps

1. **[Step name]**
   - Create/modify: `path/to/file.ts`
   - Details: [what specifically to do]
   - Verify: [how to confirm this step works]

2. **[Step name]**
   - Depends on: Step 1
   - Create/modify: `path/to/file.ts`
   - Verify: [test command or check]

### Risks
- [Risk]: [Mitigation]

Executing Plans

1. Execute one step at a time
2. Verify after each step (run tests, check build)
3. Don't skip ahead -- dependencies matter
4. If a step fails, fix before proceeding
5. Update the plan if reality diverges from it

Interview-Then-Execute

When requirements are unclear, have the agent interview you before coding.

Pattern

1. "I need [vague feature]. Interview me to understand the requirements."
2. Agent asks 5-10 targeted clarifying questions
3. You answer, refining scope and constraints
4. Agent produces a plan based on answers
5. Review, adjust, then execute in a fresh session

Why Separate Sessions

• Planning session fills context with requirements discussion
• Execution session gets clean context with only the finalized plan
• Produces better code than doing both in one session

Prototype Over Spec

For exploratory work where requirements are uncertain, build instead of specifying.

Pattern

Instead of: Write a 10-page PRD → review → implement → discover it's wrong
Do:         Build rough prototype → iterate 5-10 times → extract spec from what works

When to Prototype

• New product ideas (you don't know what you want yet)
• UI/UX exploration (seeing beats describing)
• API design (try using it before specifying it)
• Algorithm selection (benchmark, don't theorize)

When NOT to Prototype

• Well-defined bug fixes (just fix it)
• API contracts between teams (spec first)
• Database migrations (plan carefully, prototyping is dangerous)
• Security-critical code (spec and review first)

Plan Review

Before executing a non-trivial plan, consider second-opinion review.

Self-Review Checklist

• Are steps in correct dependency order?
• Does each step have clear verification?
• Are there any missing steps?
• Is scope creep controlled (not adding extras)?
• Can each step be independently tested?

Second Opinion Pattern

1. Write the plan
2. Start a fresh session
3. "Review this plan as a staff engineer. What's missing, what could go wrong?"
4. Incorporate feedback
5. Execute in yet another session (clean context)

Anti-Patterns

Anti-Pattern	Problem	Do Instead
Planning small changes	Overhead exceeds benefit	Just do it for <3 files
No plan for large changes	Miss dependencies, ordering issues	Plan for 5+ files
Planning and executing in one session	Planning context pollutes execution	Separate sessions
Over-specifying plans	Plans become rigid, resist reality	High-level steps, details during execution
Ignoring plan during execution	Plan was wasted effort	Follow it or explicitly update it
No verification steps in plan	Can't tell if steps succeeded	Every step needs a "verify" section