Here’s number 1. No email needed.
1. It says “done.” Nothing was done.
Fix: open the thing it made. Load the page, read the file, look at the sent folder. Its report is a claim, not proof.
That one costs nothing and catches more than you’d expect. The other 12 are in the checklist. Get the other 12
Why it exists
Over 8 days my agents went through seven billion tokens. 96% of it was the same context being read again and again. That wasn't the AI being bad. That was my setup. These 13 are what broke on it, and what I changed.
What's inside
- Why “zero errors” can be the worst reading your setup gives you, and the second check that makes it honest
- The quiet log that looks exactly like “nothing to report”, and the one line that tells a dead job from a quiet day
- How a one-word search flagged two unrelated videos as repeats, and the fix
- Why the instructions file you keep adding to gets more expensive on every single call
- What to do when your agent hits a blocker: the difference between a route around a broken tool and a route around the truth
- A rules file you paste into Claude or ChatGPT that puts eight of the 13 fixes straight in (10 rules)
Who it's for
Anyone using an AI agent for real work who'd rather it just worked than follow AI news.
Not for you if you want model news or prompt tricks. There’s none in it.