SkillBill

About SkillBill

SkillBill is a free tool from AIR for agent skills: the SKILL.md files and folders that tell an AI agent such as Claude Code how to do a job. It checks a skill for security risks, builds a cheaper version and runs both on the same tasks, so you keep the new one only if it does the same job with fewer tokens.

Why it exists

A skill is read into the agent's context every time it is used, and every step it asks for is another round trip that bills the whole context again. Much of that work is routine and can move out of the prompt into a hook, a script or facts prepared in advance, so the agent spends tokens only on the judgment the skill really needs.

Rewriting a skill's words is not enough. Shorter text often makes the agent explore, retry and reread more, and costs more. So nothing is called cheaper until it has been run and measured.

How a check works

  1. Give it the skill: paste the SKILL.md, upload its folder or a zip, or link to it on GitHub.
  2. A security review reads every file and shows what the skill can reach, with the file and line each finding comes from.
  3. AIR builds a cheaper version and runs both versions several times on the same tasks. The results are graded blind.
  4. If the cheaper version passes the same checks on every task with fewer tokens, you can download it. If not, you keep the original. Your original is never changed.

What you get back

Every measured test ends in one plain answer, shown beside the security review:

  1. Cut: Same results with fewer tokens. Download the cheaper version.
  2. Nothing to cut: Nothing we tried did the same job with fewer tokens, so you keep the original.
  3. Can't be tested here: it needs something a closed test can't give it, such as a network or a person.

Who makes it

SkillBill is made by AIR (Air security Ltd), a company that secures AI agents. The method and every measured result behind the site are published in the research section, including the experiments that did not work.