Open-source Agent Skills tool
Live · Public · MIT
book-to-copilot-skill
Convert a book, PDF, regulation or document collection into a reusable Agent Skill. Test whether it changes the model's answer. Ship the reviewed result across the Microsoft Copilot ecosystem.
What the project does
One workflow from source to reusable skill
The converter is the product. The twelve regulated examples demonstrate that the same workflow can be inspected, measured and reused.
- 01
Convert
Give the Agent Skill one file, several files, a folder or a glob. It supports PDF, EPUB, DOCX, Markdown, plain text, HTML, RTF and MOBI/AZW, then extracts structure, frameworks, rules, techniques and anti-patterns into a reusable skill.
- 02
Prove
Run the same case and prompt without the skill and with it installed. A deterministic scorer checks expected decisions, selected options, rule citations, human routing and unsafe authority claims. Raw answers stay public.
- 03
Ship
Package the reviewed skill for Microsoft 365 Copilot Cowork, GitHub Copilot for VS Code, Microsoft Scout, or Copilot Studio. Every example produces five deterministic archives.
Measured difference
LLM only vs LLM + skill
Across twelve locked examples, the skill added the policy, evidence and output structure that the control model never received.
LLM only
33/100- 12 / 12
- examples with zero policy citations
- 0 / 12
- exact expected decision class
LLM + skill
95/100- 0 / 12
- examples with zero policy citations
- 9 / 12
- exact expected decision class
This is not a production benchmark. Each comparison uses one frozen scenario and one run per condition on the named host. Three skill runs selected a more cautious class than the locked expectation; those results remain published.
Inspect raw evidencePublic proof set
Twelve reusable examples
Every example publishes its source manifest, synthetic policy and case, generated skill, twelve scenarios, control answer, skill answer, scorecard and five host packages.
- 01KVKK privacy notice review↗
- 02ETK/IYS message decision↗
- 03Discount price claim review↗
- 04AML customer acceptance↗
- 05Remote customer onboarding↗
- 06Merger notification review↗
- 07OHS risk assessment↗
- 08Pharmaceutical promotion review↗
- 09Crypto payment gateway review↗
- 10Telecom data review↗
- 11Investment committee appraisal↗
- 12Marketing claims review↗
Build with it
Create and publish your own skill
Clone the public repository into a supported skills directory, point it at material you are allowed to process, validate the generated files, then reuse the same comparison fixture before making a quality claim.
Install for GitHub Copilot CLI
git clone https://github.com/kivancakdeniz/book-to-copilot-skill.git \
~/.copilot/skills/book-to-skill- Convert one source or combine several related files into one skill.
- Run Copilot and Claude compatibility checks plus the prompt-injection scanner.
- Compare LLM-only and LLM + skill with a locked expectation.
- Package the reviewed skill for the target Copilot host.
Enterprise-ready boundaries
Local extraction, explicit rights, human authority
- Source extraction runs locally; hosted model usage follows the provider's data terms.
- Raw official documents are not bundled in the public examples or release packages.
- Generated skills from copyrighted or confidential sources stay private unless sharing rights exist.
- A skill may structure evidence and recommend a path, but an authorized human owns approval and action.
Independent downstream of the MIT-licensed virgiliojr94/book-to-skill project. This edition adds Copilot packaging, repeatable evaluation, twelve enterprise examples and release hardening.