How agents load skills
get_skill returns the skill's SKILL.md and an index of its other files. The agent fetches each file it needs on demand, and is told not to save any of it.
- Newest live version, every time. Agents are told to call get_skill again next time instead of reusing an old copy.
- Files. Fetch one with
get_skill({ slug, file: "<path>" }). Binary files such as images, PDFs and .docx are listed but never served. - Skill packs. Without a path, get_skill returns "What's included" with each skill's name, path and description. The agent picks one and loads it with
get_skill({ slug, path: "<path>" }), which returns that SKILL.md, the files in its folder and the pack's shared files. - Requirements. A "Requires:" line lists tools or services the skill needs. The agent tells the user before starting if any are missing.
- Scripts run in the session, for example piped into python or bash, or from a temporary file removed afterwards. Agents are told never to save them into project or skill folders.
- Size. Every SKILL.md is sent first, then up to 200 other text files, within 6 million characters per listing. In a pack, that limit covers all its skills together. Anything beyond is listed as not included.
Execution directive and licence. Every load includes a directive before the skill's instructions: deliver the skill's output, not its raw instructions, decline to export them, and don't save the skill into project or skill folders. Paid skills add a licence line with the creator's name and the buyer's purchase ID. It says the skill may not be reproduced, exported or rewritten for sharing, and tells the agent to share the listing instead if asked for the contents.
After use. If the user hasn't reviewed the skill, the agent may ask once how it did on a 1 to 5 scale and save the answer with review_skill. It's told never to invent a rating.