AI-Assisted Planning
Jump to section
Why Plan with AI
Most developers start using AI only at the coding stage. But the biggest time savings come from planning — AI helps you break features into tasks, identify edge cases upfront, and create an implementation plan before you write your first line of code.
What AI Does Well in Planning
- Decomposing features into concrete implementation steps
- Identifying dependencies between tasks
- Designing database schemas and API contracts
- Reviewing architecture decisions — finding gaps in the design
- Generating acceptance criteria from vague requirements
Context Is Everything
AI can only plan well what it understands. The key is providing context: existing architecture, technical constraints, business requirements. Without context, you get a generic plan that ignores the reality of your project.
CLAUDE.md and Project Context Files
CLAUDE.md is a file at the root of your repository that provides AI agents with project context. It contains your stack, conventions, directory structure, workflow — everything a new developer would need to know. Agents like Claude Code read it automatically.
# CLAUDE.md — example structure
## Stack
- Next.js 16, TypeScript, Tailwind CSS 4
- PostgreSQL + Prisma ORM
- Vercel deployment
## Conventions
- Components in src/components/, pages in src/app/
- Tests alongside files: Component.test.tsx
- Commits: conventional commits (feat:, fix:, chore:)
## Workflow
- PR to main, CI must pass
- E2E tests in PlaywrightPrompt Patterns for Planning
# Feature decomposition
'Break this feature into implementation tasks:
[feature description]
For each task include: files to modify, dependencies, estimated complexity.'
# Architecture review
'Review this architecture design. Look for:
- Single points of failure
- Scalability issues
- Missing error handling
- Security risks
[architecture design]'
# Complexity estimation
'Based on this task list, estimate relative complexity (S/M/L/XL).
Consider: number of files, dependencies, test coverage, regression risk.'You get the best results iteratively. Start with a high-level plan, then ask AI for details on individual steps. Do not try to get a complete plan from a single prompt.
Example: From Feature Request to Implementation Plan
Imagine a feature request: 'Add user data export to CSV.' Instead of coding immediately, use AI to create a plan:
# Step 1: Define scope
'Feature request: Export user data to CSV.
Our stack: Next.js API routes, Prisma, PostgreSQL.
What should the export include? What edge cases to consider?'
# Step 2: Technical design
'Based on the previous analysis, propose a technical implementation.
Consider: streaming for large datasets, authorization, data format, limits.'
# Step 3: Task breakdown
'Break the implementation into PR-ready tasks.
Each task = 1 PR, max 200 lines of changes.'AI tends to plan for greenfield projects. Always specify existing constraints: 'We have an existing export module at src/lib/export.ts — extend it, do not create a new one.'
Explicitly tell AI about existing constraints: 'We have 2 developers, sprint is 2 weeks, tests must pass in CI within 10 minutes.' Without constraints AI designs ideal solutions, not pragmatic ones.
Take a real feature request from your backlog (or use: 'Add email notifications when order status changes'). Using AI, create: 1) Scope analysis and edge cases, 2) Technical design considering your stack, 3) List of implementation tasks ordered by dependencies.
Hint
Give AI maximum context about your existing code — what models you have, what email provider you use, where existing notification logic lives.
Use AI as a planning assistant at your next sprint planning. Paste backlog items into AI and ask: 'Here are user stories for next sprint: [stories]. Help me: 1) Estimate complexity (S/M/L), 2) Identify dependencies, 3) Suggest implementation order, 4) Identify risks. Context: team of [X] devs, [Y]-day sprint.' Compare AI estimates with team estimates.
Hint
Document your process and results — they'll serve as reference for similar future tasks.
Pick a medium-sized feature (1-2 weeks of work). Prompt AI: 'Write a technical design doc for: [feature description]. Include: goal, proposed solution, alternatives and why they were rejected, API contract, database changes, migration strategy, test plan, risks.' Discuss the result with a colleague — how much of the design doc could you use directly?
Hint
Add your existing CLAUDE.md or project README as context — AI will then propose solutions compatible with your architecture.
- AI in planning saves more time than in coding — it identifies problems upfront
- CLAUDE.md and context files are the foundation of quality AI planning
- Plan iteratively: scope -> technical design -> task breakdown
- CLAUDE.md is a key input for quality planning — AI needs to know your stack, conventions, and constraints
- AI design docs save hours of work but always need human review for business context
In the next lesson, we dive into Coding with AI — a technique that gives you a clear edge. Unlock the full course and continue now.
2/8 complete — keep going!