SkillBill

Context Catchup

FlorianBruniaux/claude-code-ultimate-guideSource at 73e1dec (opens GitHub)

Nothing to cut.

We built a cheaper version and ran both on the same tasks. It wasn't better, so the original stays as it is.

  • We tried prepared facts, and didn't keep it because the saving wasn't consistent across runs.

What AIR found

  • Not checked yetSecurity review

    AIR hasn't reviewed this skill yet. Check it to run the review.

On a smaller model

Same results on a smaller model, for less.

34% cheaper per run

A separate result from the token saving: this is the cost of running the same skill on a smaller model, and the two are never added together.

Each task on the smaller model against the default model
TaskResultsCost
Context restoration after breakSame results34% cheaper
Prioritize work from stateNot comparedNot compared

Measured in the prepared-facts study, Sep 26, 3 runs per task, graded blind. How we measured

Embed badge
SkillBill result
Markdown
[![SkillBill result](/badge/catchup.svg)](/skills/catchup)

HTML
<a href="/skills/catchup"><img src="/badge/catchup.svg" alt="SkillBill result" /></a>