Shortening an AI skill can cut the exact details that give it value

A skill file is a set of written instructions that an AI agent loads to do a task well. Nikolay Petrov of Intel argued that shrinking one by instinct is risky. The model already knows the general idea, he said, but not the exact details. A shortener that cuts what looks obvious may remove the part that carries the value. He made the case on the AI Engineer show.

What was said

Petrov described a way to look inside a skill. Split it into small claims that can be checked. Then test the model with and without each claim, to see what the skill teaches and what the model already knew.

He ran two kinds of check. One asked whether the model grasped the general idea. The other asked about exact keys, parameters, values and steps. On larger frontier models, he said, the gap was big. The result was "dropping from 70% understanding of their skill in case of the general understanding to just 22% of the specific details".

The show notes add that the first skill tested was 70% things the model already knew. The value sat in the other 30%, the details a naive shortener would cut first. Petrov said the run was cheap: about 14 dollars of model calls for 29 skills.

Why it matters

Our reading: this moves the work from writing skills to judging them. Adding a line takes seconds. Deciding which line to remove needs evidence, not taste.

For teams that build or buy agent tooling, the lesson is plain. A skill that looks padded may be mostly background the model has, plus a small core of specifics it lacks. Cutting by length or by what sounds generic could keep the filler and delete the substance.

Petrov offered a simple path. Remove the parts the model already knows, run the evaluation, and see if the score holds. If it does, the skill gets smaller and saves tokens, which are the units of text a model reads and bills for. In his words, this replaces brute forcing and guessing with guidance.

The other side

Petrov was careful about what the method proves. He called the measurement "basically initial vector for optimization at the end we still have to run evals and validate". It is a signal, not a verdict.

He also noted that results differ a lot from skill to skill. Some skills are meant to be directional, and low novelty there is not a flaw. Smaller models understood less of a skill than larger ones, so the right cuts may depend on which model runs it. He said the test should ideally cover a broader set of models.

The talk did not say how often a trimmed skill kept its score in real use, or how the figures hold beyond the skills Petrov tested.