Thesis
The cost of a skill library is routing attention, not tokens.
On a production agent with about 100 skills, the always-on skill block was about 11,500 tokens per turn, but the prompt cache hit rate was 87 percent, so real spend was roughly a fifth of nominal. Token cost was never the constraint.
01 · Argument
Treat the skill block as a routing table, not a library.
Every description is a candidate answer in a multiple-choice question the model answers on every turn. Adding a skill does not only add tokens. It adds a wrong answer. Prune for discriminability, not size.
Evidence
Descriptions are uniform, 300 to 374 characters, so only fewer skills helps.
Median 332, minimum 300, maximum 374 across 100 skills: there are no outliers to trim. Eighteen percent of skills had never been invoked, and a third had been invoked exactly once.
Diagram · Split by consumer
Preconditions go to the runtime as structured data and never reach the prompt. The semantic trigger goes to the model as natural language, and only when the preconditions pass.
“Adding a skill does not only add tokens. It adds a wrong answer.”
“A skill that is never invoked is worse than unused. It is an active wrong answer in every routing decision.”
Tension
The headline saving was imaginary.
An audit found 62 percent of skills declared a gateable dependency and projected a 44 percent saving. The platforms in use shared nearly identical toolsets, so almost no gate could fire. The pack records the error as an anti-pattern instead of dropping it.
Implication
Demote, do not only kill.
The awkward cohort is skills invoked exactly once over months. Killing them loses capability and leaving them taxes every turn. Demotion keeps them reachable by name and out of the routing block.
Provenance
Derived from an audit of a production agent's routing block, including the numbers that overturned the audit's own first conclusions. The raw audit table is not published with the pack.