The reason why Claude Code's skill is "chosen but not read": measured by dividing activation into two stages

It was reported that whether a Claude Code skill is used is decided in two stages. The first stage is "whether it enters the candidates by looking at the description," and the second stage is "whether the model actually reads that skill on the spot."

In a test where only the description list was shown to let it choose, 9 out of 9 chosen skills were never read once in actual work. Even strengthening the description did not change this, and the skill was read for the first time only when inserted with "read this skill" via the UserPromptSubmit hook.