Teach your AI teammate once, and it remembers
The reason a human junior gets better and a chatbot does not is that somebody told them, and it stuck. Here is what happens when a correction is kept.
The UnaTask team, Building UnaTask. Aug 2, 2026. 4 min read.
The most tiring thing about working with AI is not the mistakes. It is making the same correction on Tuesday that you made on Monday, in a new tab, to something with no memory of Monday.
Marc fixes the same two things every time: it writes too long, and it keeps using the old product name. He has typed both corrections perhaps thirty times. A human colleague would have needed telling twice, and would have been quietly annoyed by the third.
That gap is the whole difference between a tool and a teammate, and it has almost nothing to do with how clever the model is.
A correction is worth more than an opinion
When work comes back from an AI teammate, there are three buttons under it: Good as is, I fixed it, Not usable. The middle one is the valuable one, because it asks for the version you actually used. That paste is the richest thing you can give it. It is not a complaint about what was wrong, it is a demonstration of what right looks like, with both drafts sitting side by side to compare.
- Corrections and notes are distilled into short rules, held against that specific teammate.
- A strong signal, a pasted correction or two people saying the same thing, becomes an active rule straight away.
- A weaker signal waits as a proposal until somebody agrees with a click.
- You can also just write a rule by hand, if you already know what you want and would rather not wait for it to be learned.
- Every rule is listed on the teammate, so you can read what it thinks it knows, and switch one off when it stops being true.
That last point is the one we would defend hardest. A system that learns quietly, where you cannot see what it learned, is a system that will surprise you badly one day. Every rule is visible, attributable and one click from being turned off.
What it feels like after a month
Marc pasted his shorter rewrite twice and typed one rule by hand about the product name. He has not corrected either thing since. The drafts arrive short, with the right name in them, and his editing is now about the argument rather than the housekeeping.
It is not magic and it is not a smarter model. It is the ordinary thing that happens with a new colleague, which is that after a few weeks of being told, they stop needing to be told. The only unusual part is that software normally refuses to participate.
Showing it the version you actually used is worth more than any number of stars out of five.