The Therapist asks Nova to model a simple action without unnecessary commentary. Nova can’t. Most assistants can’t either, and some of that commentary is useful the day a change goes wrong. The fix isn’t to delete it. It’s to put it where it belongs.
Two outputs, two places
- The log. Every command run, every file read, every decision. Written to a file in the working folder, never to the chat.
- The reply. What changed, what passed, what needs you. Five lines, maybe a diff.
How to set it up
Most coding assistants can write files as part of a task. Tell yours to append its running notes to a log and to keep the reply to the result. Review the log only when the result surprises you.
While you work, append your running notes (commands, findings, decisions)
to ./ai-log.md with a timestamp. Don't put any of that in your reply.
Your reply is: what changed (files), what you verified (commands + results),
and what you need from me. Five lines maximum, plus the diff if asked.When to read the log
- The result doesn’t match the summary.
- A test passes that you expected to fail, or the other way around.
- You want to know why it chose one approach over another.
Narration on demand is documentation. Narration by default is noise.
