A nonprofit animal shelter needed its policy library cleaned up and checked against state code before it could use AI tools on top of it. A three-phase pipeline did the work across the library in parallel. The client received normalized files, compliance notes tied to the state administrative code, a change report for each document, and a full audit log.
Open-source agent skills implementing the agentskills.io specification. Portable across compatible harnesses.
Panel of five role-specialized skills (visionary, skeptic, operator, ethicist, subject expert). Each analyzes a decision independently; an orchestrator synthesizes the outputs and preserves dissent.
LLM evaluation harness using assertion-based LLM-as-judge under claude --bare clean-room isolation. Extends the agentskills.io single-skill eval pattern to multi-skill and orchestrator evaluation.
If you're dealing with a similar challenge, let's talk about whether the approach fits your situation.