You are improving a set of instructions (a skill) for an AI. A cheap model takes this skill together with a list of names and terms, and corrects speech-to-text transcripts one at a time: it should fix names and terms that were transcribed wrong and leave everything else as it is.

We score it on an evaluation set: in each transcript, every name or term written correctly earns a point, and every name or term added that is not in the original (a wrong insertion) loses a point. Score = (correct − wrong insertions) / total that should be correct.

The current skill:
<skill>
{skill}
</skill>

The list of names and terms (it is given to the model together with the skill; you cannot change it):
<names>
{names}
</names>

The current skill's score on the evaluation set: {score}

{history}
Below are the items in the evaluation set that went wrong, one JSON object per line: raw is the raw transcript, output is the model's output, expected is the correct answer, missed lists the words that should have been written correctly and how many times, wrongly_added lists the words that were inserted and how many times.

{failures}

Rewrite the skill so the score is as high as possible. Output only the full text of the new skill, with no explanation and no code fences.