Notes on working with AI
Small tips, tricks and ideas I use with Claude and other AI tools. I'm not a coder, but more experienced than a 'Vibe coder'. I'm comfortable in a command line and know generally how software and data fit together. I write for people like me.
- Vibe Coding Built a Hacker's Paradise
Anyone with a keyboard can ship an app now. That also means anyone with a keyboard can go looking for a way in. Be careful what you store.
- Number Your AI's Replies
Long AI chats turn into a scavenger hunt for who said what. One line in your custom instructions numbers every reply so you can point at exactly the one you mean.
- You're Probably Vibe Coding in Production
Most vibe coders are testing changes directly on their live app and database. A cheap habit from 20 years of enterprise IT fixes that, no IT department required.
- Let AI's Review Each Other's Work
When I started building my investing engines, I stopped trusting one AI's first draft. Here's how I use a second (or third, or fourth) AI to catch what the first one missed.
- From Hard Stops to Silent Drift: What I've Learned About AI's Memory Problem
Context windows either hit a hard stop or quietly drift — here's how I keep AI from silently forgetting the facts that matter, in Claude Code and in browser chat.
- Stop AI from Guessing
I noticed my Claude Code guessing at column names during its 'thinking'. This wastes tokens, time, and is a technical risk. This is how I resolved the guessing.
- Hit the brakes at 60% context
I end most Claude Code sessions with more than half the context window unused. Keeps Claude sharp, quality high, and saves tokens.
- End every AI session with a /handoff
The single habit that made long projects with Claude survivable: the AI writes its own note to its future self before the session ends.