Useful Prompts

This is a collection of useful prompts and prompt ideas that I like. These are patterns and techniques I’ve found effective when working with AI assistants like ChatGPT and Claude.


Capture Your Favorite Conversational Style

Wait until you get a session with ChatGPT or Claude that you’re really enjoying. Maybe you’ve been joking around, maybe the AI has gotten introspective and empathetic and you like that in a conversation partner. At that point, use this prompt:

“Summarize your own voice in this conversation—your tone, style, pacing, and rhetorical habits—so I can copy it into my Custom Instructions. Leave out any references to specific content. Just describe the stylistic traits I’d get if I wanted future chats to feel like this. Write it as if you’re an introspective writer describing their own persona.”

Then copy-paste the results into your Custom Instructions (in Settings). Now you have an AI assistant that you LOVE to talk to.


Tune Up Your Mac With Claude Code

Open Claude Code on your Mac and run this:

“Look in /var/log at my system logs and make suggestions on how to improve my Mac’s performance. There’s obviously a LOT of data in there, so use good judgment to get enough signal without processing everything — e.g. awk, head, tail, don’t unarchive old logs.”

Then go make a coffee. Two minutes later you’ll come back to a report.

When I ran this, it found an old version of MySQL crashing in a loop, SOC2 monitoring agents from past jobs still phoning home, out-of-date software, a network misconfiguration, and a bunch of other detritus my laptop had accumulated over the years. Plus remediation steps for each one.

This isn’t a human-tractable problem. No sane person is grepping /var/log on a Saturday. But it’s basically a unit test for Opus, and it really does breathe new life into a crufty laptop.

Let me know what you find, and whether it fixes your [running out of memory / slow wifi / fan won’t stop / battery dies fast] problem.

(Disclaimer: don’t blindly run sudo commands that an LLM tells you to. I mean, I do. But you shouldn’t. Or at least don’t blame me.)