When a smaller model fails at a task your agent repeats, the reflex is to train it.
Training takes weeks, and it often puts into weights what one page of instructions could say. The gap is usually unwritten context: which tools to use, the format, and what to do when a step fails. A top model works that out on its own, but a smaller model needs it spelled out.
Operant writes that context down first. For each repeated task, it learns a skill from your successful past chats. A skill is a short file that states the tools and the judgment the task takes. You can read it, edit it, export it and keep versions of it. It also records which chats taught it.
The skill isn't a copy
In our published run, the skill started from two starter files that held only method and format. Together they were 6,155 bytes. The learned skill was 7,270 bytes, more than both starter files combined. So it can't just be the starter files stitched together.
It cleared the bar without training
Then a smaller model, carrying that skill, answered chats it had never seen. It was graded against the top model's own answers on those chats.
| What | Value |
|---|---|
| Starter files, method and format only | 6,155 bytes |
| Learned skill | 7,270 bytes |
| Score on chats the skill never saw | 0.9667 |
| Bar to pass | 0.8 |
| Weights changed | None |
Why the order matters
Writing the skill first is cheaper, and it also tells you something. If the skill closes the gap, the gap was context. Training would have spent weeks putting into weights what a page of text says directly.
If the skill doesn't close the gap, you've learned where the model really falls short. Any later training then has a clear target and a ready pass mark: the same test the skill just failed.
There's also a control argument. You can read a skill before it's used, and you can see what changed between versions. You can't do that with weights. When a silent drop in quality costs real money, start with the change you can read.
In that run, the gap was context, not ability. A page of instructions closed it, and no weights changed at any point.
So you only pay for training once a written skill has shown it isn't enough.
See how a skill gets written in Teaching a cheaper model how your team does a task.