On this page
I asked Claude Code to go through my own AI conversation logs from the past month, more than 300 sessions, and find habits I keep using that a lot of people around me don’t seem to know about. Later I used the same logs for two more things.
Want AI working in your own job, not just in demos? See the 3-month 1-on-1 program →
Four habits
I picked four and made them into a short video:
- When the AI says “the email’s sent,” I make it dig up the sent record and show me. If it can’t find one, it has to say so, not guess.
- The AI that writes and the AI that checks are separate. The second one finds problems, then I hand those back to the first one to verify item by item, and it only changes what holds up.
- Whatever it can decide on its own, it just does. Anything that needs my call waits until the end, as a numbered list with recommendations, and I just reply with numbers.
- When it says “I’ll be more careful next time,” I don’t accept that. I make it write the new approach into the project instructions and paste it back for me to see.
While digging through the logs it also found me telling the AI, “Whatever you say about later is just talk, right?” Reading it now, that was a bit harsh XD
Here’s the short: https://youtube.com/shorts/5g1jYJ3DZRE
Turning catchphrases into switchable Mods
I turned the three things I say to AI most often (don’t ask me, delegate more, let’s wrap up) into switchable auto / out / wrap modes. I later wrote it up in full, with a week of measurements and the prompts I used to build it: Switching Your AI’s Working Mode with One Command.
The two other Mods I built earlier are here: a recording mask and a warm-cache plugin.
Auto mode keeps blocking you? Dig through the logs too
Someone complained that auto mode’s automatic review kept blocking them. That used to annoy me too. My suggestion: have it read the past month of session log jsonl, find every case where the automatic review refused something, and rewrite your auto mode rules from that. Apply them by hand once, and you’ll rarely hit the problem again.