Skip to content

Ian Nuttall's Tips for Using Claude Code

Source: Ian Nuttall's Tips for Using Claude Code (also mostly works for Cursor + other coding tools)

1. Take longer than you think to plan

  • Use Claude Desktop in open-ended conversation to discuss:
    • Tech stack
    • Packages and libraries to use
    • MVP features needed
  • Verify suggestions manually, especially for unfamiliar stacks.

2. Find URLs and repos to share

  • Claude might not know the latest packages.
  • Provide URLs to documentation and GitHub repos.
  • Pro tip: Use uithub.com instead of github.com for LLM-friendly repo context.

3. Ask Claude to write a highly detailed, step-by-step plan

  • Request a meticulous, executable plan that any human or agent can follow.
  • IMPORTANT: Don't give Claude Code the entire plan upfront. Work through the plan one step at a time as the conductor.

4. Get a working version ASAP

  • Don't try to build the full app at once.
  • Focus on getting a bare minimum MVP working first.
  • Iterate from there. It's easier to add 1 feature than 10 at once.

5. Run /init in Claude Code to create a CLAUDE.md file

  • Do this after the MVP is working.
  • Use # comments to add memories later, e.g.: # Always run bun typecheck to check for type errors
  • Root CLAUDE.md should contain project-level rules only.

6. Add CLAUDE.md files to subdirectories

  • Create per-directory rule files:
    • /src/components/CLAUDE.md
    • /src/db/CLAUDE.md
  • Claude uses this for precise context. It makes a HUGE difference.

7. Track every session

  • Log each session by answering:
    • What are you working on?
    • What did you do in this session?
    • What worked? What didn't?
    • What are the next steps?
  • Use claude-sessions for memory and tracking.

8. Use /clear as early as possible

  • You can use /compact, but detailed session logs are better for preserving context.
  • Clear as soon as you hit a good stopping point.
  • Claude becomes less effective after the context warning.

9. Use subagents and "ultrathink" sparingly

  • Powerful tools, but they burn context quickly, making Claude feel "dumber".
  • Use them for understanding and planning.
  • Then save the session and /clear before continuing.

10. Use Repo Prompt for tough bugs/features

  • o3 is a great coding model but not a good agent.
  • Use Repo Prompt MCP to let Claude:
    • Find context
    • Pass task to o3
    • Get it implemented

11. Use Playwright to test your app

  • Playwright can:
    • View, click, navigate, use console
    • Test your app
    • Send back info to Claude to fix bugs
  • Add usage instructions to CLAUDE.md.
  • Run this command from your terminal:
    claude mcp add playwright npx '@playwright/mcp@latest'
    
  • Restart CC, then you can ask things like "use playwright mcp to test the button on my homepage"

12. GO SLOWER!

  • Don't rush by launching 6 agents and 20+ subagents.
  • You'll lose track of your context and codebase.
  • For deep work, slow down. Understanding your code is 100x better.