AuditHQ
← All articles
Article · Claude plugins · 8 min read · Updated 2026-09-03

Claude Skill Context Cost: Progressive Disclosure That Actually Saves Tokens

Standing context is a product tax. Pay it only when the skill is selected — and only for files that run needs.

Standing cost vs on-selection cost

Descriptions and selection signals are cheap compared with a skill body that forces large supporting files into every turn. Progressive disclosure means: keep the procedure in SKILL.md, and load deep references only when that run needs them.

If a skill always reads a long shared runtime file before checking whether the user already named the artifact, you pay tokens on work the user already finished.

What belongs upfront

When the skill applies. The smallest procedure that produces a correct result. Guarded early exits when the user already named the file or command. Explicit “leave untouched” boundaries so edits stay additive.

What does not belong upfront: encyclopedias of edge cases, historical changelogs, and tool manuals that only matter for rare failure paths.

How to measure improvement

Compare token use and time-to-answer on the same three tasks before and after the split. Rank fixes by measured cost and the file that carries it — not by generic best-practice slogans.

AuditHQ’s Claude plugin scorecard surfaces standing context and ranked fixes so authors ship the expensive cut first. One company using that backlog cut token spend about 40% and sped skill runs about 50%.

Do not confuse hygiene with cost

Naming, keywords, and attribution matter for discovery. They are not the same band as unconditional per-run context. Fix the meter first; polish the listing second.

Score your Claude plugins on a schedule

AuditHQ tracks Claude plugin listing quality, peers, and ranked fixes — so ratings stay current after each Claude Code release.

Claude skill context costprogressive disclosure SKILL.mdreduce Claude Code tokensClaude skill upfront readsSKILL.md token wastefaster Claude skills