Tips for Claude Code

Must Read or Watch

• Claude Code & the evolution of agentic coding — Boris Cherny, Anthropic
• A conversation on Claude Code — Boris Cherny, Anthropic
• Claude Code: Best practices for agentic coding — Boris Cherny, Anthropic

Use Every Feature

Use every feature of Claude Code—not doing so is leaving power on the table.

• Use dictation to give the model more context; every word helps narrow down the task to what you were expecting.
• Use the plan mode.
• Use ctrl+z to exit Claude Code to terminal, then resume with fg.
  • It's so much better than the ! bash command for running commands WITHOUT adding them to the context.
• Use the ! bash command to run commands that you want added to the context.
  • Running --help on a tool is a great way to get the model to understand the tool.
• Use the memory.
• Use the web search.
  • I've been linking to the README of the repo I'm using at the time to load it into the context.
• Use up arrow to repeat commands.
• Add commands to .claude/commands to make them available in the future and with $ARGUMENTS to pass arguments to the command.

Don't Get Distracted

• If you have an idea how to improve something, create an issue to remember it, but don't get distracted by it in the moment.

Other Tips

• Think about how you manage context. The idea may be already now that the best communicator → the best coder. The New Code — Sean Grove, OpenAI
• So writing document specifications may be even more important than writing code.
  • Give a look at these requirements for a project, which really embraced this idea. I used Claude to translate from Japanese to English: Requirements for a Project
• Use pictures.
  • Claude Code can use pictures, so if you have a picture of the code you want to change, you can use that to give it more context.
  • Give Claude access to output of what running the code does so it can create a feedback loop to really improve.
    • Tests are great for this, but you can also use screenshots or text of the output.
    • For any web app, you can use Puppeteer to take screenshots of the output and use that as context.