Turn Documents Into Runnable Skills

Documents are incomplete specifications with unstated assumptions about who runs them, what tools are available, how to handle missing inputs, and whether the described process actually works. Doc-to-skill extracts what the document actually says, interviews the creator for what it doesn't, generates a complete skill with QUALIFY/SETUP/RUN modes, and validates it with a realistic simulated run so that hidden assumptions surface before the skill ships.

1. Provide any document (playbook, SOP, deck, rough notes) that describes a repeatable process with a concrete output. 2. The skill ingests the document, builds an extraction sheet, and conducts a guided gap interview using discovery (checking your connected tools and data first, then asking only for genuine unknowns). 3. It generates a complete skill folder with SKILL.md (agent instructions), README.md (human docs), and config.yaml (all tunable parameters). 4. It runs a simulated client execution end-to-end using real tool reads but no real writes, detects the classic five defects (missing capabilities, query traps, rules that don't fit real data, silent failures, fabrication risks), and shows you the sample output for feedback. Output: a tested, documented, configurable skill ready to evaluate.

The skill is agent-agnostic and uses plain-language instructions, so it runs in any AI assistant with tool-calling and file read/write access. Simulated runs use real API calls for reads (authenticating with your connected tools) but mock writes with clearly labeled [SIMULATED] tags to prevent unintended changes during testing. If tools are unavailable or behind authentication, the skill uses obviously fictional test data. No API key is needed to run the skill itself, but the skills it generates will likely need credentials for the tools they automate.