Skip to content

Safety & reuse

Use this project with material you are allowed to process. The converter and the public examples do not grant rights in source books, regulations, internal documents, screenshots, or model output.

Source rights

  • The repository ships no source book or complete official document.
  • Keep generated skills from copyrighted or confidential sources private unless you have permission to share them.
  • Official sources in the 12 examples are represented by metadata and short, independently written method summaries.
  • Synthetic policies, cases, evaluation fixtures, downstream code, and authored documentation are released under the repository MIT license.

Data handling

Extraction runs locally. Do not put credentials, tenant identifiers, customer data, personal data, confidential documents, or regulated records into public prompts, screenshots, traces, issues, or evidence. If the agent model runs in a hosted service, text sent to that model follows the provider's data terms.

Generated skill review

Treat every input document as untrusted. Before installing a generated skill:

  1. inspect SKILL.md and every supporting file;
  2. review commands, links, paths, and frontmatter;
  3. run host compatibility and prompt-injection scans;
  4. remove source text that should not be retained or shared;
  5. verify the current official source and company policy.

The scanners reduce risk; they do not certify a skill as safe or correct.

Evaluation claims

The published comparison uses one frozen scenario and one run per condition on the named host. It checks a machine-verifiable subset of behavior. It does not prove production accuracy, legal compliance, safety, fairness, robustness, ROI, or future performance.

Publish raw answers and scoring rules when claiming that a skill improved a model. Do not hide runs that disagree with the expected result.

Human authority

A skill can structure evidence and recommend a review path. It cannot approve, publish, submit, onboard, reject, suspend, transfer, send, stop, or execute an operational decision. An authorized human owns interpretation, approval, and action.

Project relationship

This is an independent downstream of the MIT-licensed virgiliojr94/book-to-skill project. It is not endorsed by the upstream author, Microsoft, GitHub, model providers, or any public authority.

Report security issues through the repository's private vulnerability-reporting flow, not a public issue.