Claude Code

So I've been playing around with Claude Code! I've only used it for a week or so at this point but it's replaced cursor as my main AI assistant.

Some tips (that you can also get from --help):

Command-line flags

• --dangerously-skip-permissions is the equivalent of cursor's agent mode with auto-run (on everything) turned on. Use it if you have some long-run task that you're okay with Claude working on without oversight.
• --resume is how you resume old conversations; --continue resumes the last one.

In-chat commands

• /help gives you a help list. Of note from those are:
• # at the start of a message makes the message a memory - prepended to the prompt. Use these for patterns that you want claude to hew close to.
• @ anywhere allows you to insert something in-context.

Note that claude code can read images just like any other file! Use this to place design docs in-context.

You can also migrate your mcp settings from claude desktop to claude code.

As you write lots of code and have longer and longer claude code sections, you'll eventually hit the 200k token limit. In the bottom-right corner you'll see a percentage left; when this is hit, it'll auto-summarize, taking about a minute before proceeding. Around when this warning starts to happen quality tends to lower. I recommend as much as you can keeping the context window small and putting shared conversation docs in a dedicated docs folder.

That's all the thoughts for now.