Getting Started
Ready to try this methodology on your next project? Here's the checklist to work through before your first commit.
Prerequisites
- Claude Code installed and authenticated (
npm install -g @anthropic-ai/claude-code) - A blank repo initialized with your framework of choice
- A
CLAUDE.mdfile at the project root (this is your persistent instructions file) - A clear idea of what you want to build — but resist the urge to start coding
Pre-commit checklist
Before you write your first line of application code, make sure you've completed:
- [ ] Phase 1 — Research. Fire off subagents to research your domain, existing solutions, and available APIs. Spend at least 15 minutes here.
- [ ] Phase 2 — Specification. Have an interactive session with Claude to define scope, identify risks, and document decisions.
- [ ] Phase 3 — Stories. Break the spec into user stories with acceptance criteria and a Definition of Done.
- [ ] Set up your CLAUDE.md. Include the visual QA rules and any domain-specific instructions.
- [ ] Set up Linear (optional). Connect the MCP server for project tracking.
I'm starting a new project: [describe your project in 2-3 sentences]. Before we write any code, walk me through: 1. What should we research first? 2. What questions do you have about the requirements? 3. What risks should we identify early? Let's start with research. Search for existing solutions and relevant APIs.
Your CLAUDE.md starter
Add this to your project's CLAUDE.md to encode the methodology:
# Project Instructions
## Workflow
- Research before coding. Use subagents for domain research.
- Break work into user stories with acceptance criteria.
- Ask "what would make this indispensable?" at 60-70% completion.
## Visual QA (Required)
- After any UI change, capture screenshots at desktop and mobile viewports
- VISUALLY INSPECT screenshots — programmatic checks alone are not sufficient
- Describe what you see. Flag overlaps, truncation, contrast issues.
- Never commit CSS changes without visual verification.
## Quality
- Every feature needs unit tests covering happy path and key edge cases
- Responsive: test at 1920x1080 and 393x852 minimum
- Accessible: keyboard navigable, semantic HTML, WCAG AA contrast
Common anti-patterns to avoid
Warning
Coding before researching. The #1 mistake. You'll build the wrong thing faster. Spend 15-20 minutes in Phase 1 — it saves hours.
Warning
Trusting programmatic QA alone. "All tests pass" and "the app looks right" are different statements. Always visually verify.
Warning
Skipping the spec conversation. A one-shot prompt produces a one-dimensional product. The interactive back-and-forth surfaces assumptions, risks, and edge cases that single prompts miss.
Warning
Monolithic prompts. Don't try to build everything in one message. Break work into stories and execute them sequentially. Claude Code maintains context — use that.
This guide is a living document. Content is being actively developed based on real project experience.