Attaching a skill file to every prompt in a coding session is usually a net loss.

A skill file hurts most on the easy early tasks your agent already handles correctly.

Whatever gain exists belongs to a particular skill, project, and model combination, not to the skill itself.

So treat injection as a routing decision you measure per backend, and hold the file back on early tasks until error rates rise.

Skills can help, but this paper finds the helpful cases are a minority that a single published ranking cannot pick out in advance, since per-pair effects barely transfer across models.

WebDev-Skills-Bench holds the harness, project, and decoding fixed and varies only the injected file: no skill, the stack-matched skill, an equally long irrelevant one, and slice-removed variants.

Across all 4 models, injection lowers mean Pass@2 on stack-matched pairs by 1.3 to 4.2 points while raising token cost by at least 72%.

The losses concentrate on easy early tasks, running from 4.0 to 10.7 points there, while harder tasks show no consistent effect.

That looks like retry lock-in: the skill fixes structural choices the model would otherwise vary on a second attempt, so a recoverable first mistake becomes a chain-terminating failure.

– arxiv. org/abs/2608.23067

Title: "Signal or Noise? A Benchmark Study of Agent Skills in Web Development"