Our summary
An AI coding assistant writes software, but also decides whether to run a computer instruction, keep working or start another task. Mr. Buzzoni proposes giving those small judgments to Jev while Claude keeps writing code. His examples add checks before an instruction runs or Claude finishes, plus a separate program for sorting incoming issues.
The program blocks a few exact command phrases, then asks Jev about risk. External or uncertain actions ask a person. To judge completion, it compares evidence with the goal, but uncertain answers stop with a review note. A limit of five repeat attempts can also stop unfinished work. Another program labels issues before starting Claude.
The headline prices calculate only the small model questions, not a measured total bill; writing code costs extra. Commands, the folder the assistant is working in and check output can reach the model provider. A short list of blocked phrases and a confident answer do not guarantee safety, so check these examples carefully before letting them run unattended.
Key takeaways
- Keep exact rules, permission to act and limits on repeated attempts in the program.
- Look at observable evidence before calling a goal finished.
- Compare the full cost of a task; cheaper small judgments do not prove the whole assistant is cheaper.