Claude Code 2.1.261 shipped /skill-doctor, a built-in meter for what every loaded skill costs you per turn and which ones never fire. It is the 8-28 skill audit as a live instrument.
On 8-28 you shelved 50 never-used skills and cut 14 plugins after Claude studied session transcripts by hand. That audit was a snapshot. Your weekly review still weighs every proposed skill against 'per-session context load' by estimate, and the roster has drifted since (three new skills built 8-31 alone). /skill-doctor makes the estimate a measurement you can re-run any Sunday.
Every skill in ~/.claude/skills/ and every enabled plugin contributes a description line to the prompt at the start of each turn, whether or not it fires. That is the 'context tax': tokens you pay on every message so that Claude can recognise when a skill applies. Fifty skills at a sentence each is a page of prompt you never see.
/skill-doctor lists the loaded skills, marks the ones that have not been invoked in the session history it can see, and estimates their token cost. It does not remove anything. It is diagnostic only, which is the right shape: the decision to shelve stays with you and the 8-28 rule ('reverse only on Tom's word') still governs.
In your system the numbers feed one place: the anti-bloat axes in the weekly-review skill (context load, runtime hooks, orphan risk). A skill that /skill-doctor shows as unused for 30 days and expensive is a shelve candidate; a cheap unused one can stay.
Inside any interactive Claude Code session:
/skill-doctor
# then, for a skill you decide to shelve (example name):
mv ~/.claude/skills/chart-visualization ~/.claude/skills/_shelved/
# and confirm it is gone from the next session's roster:
ls ~/.claude/skills/ | wc -lThe context tax is real and now measurable. Measure on Sundays, shelve on evidence.