/source-librarian
This is part of a series about the AI agent skills I designed, built, and maintained to cover the full lifecycle of end-user help documentation.
What it is
/source-librarian is a low-cost skill that takes in Zendesk tickets, Jira issues, GitHub code, Confluence pages, Notion pages, transcripts, Slack threads, Google Docs, Google Slides, PDFs, plain text, and other sources of product knowledge so drafting skills can find material relevant to the topic they've been given to write about.
How it works
For each new file, it reads the full content, strips boilerplate, applies tags from a 16-dimension product taxonomy, and writes a PII-free summary and drafting guidance into the file's YAML front matter. A compact master index and one index shard per workflow connects the dots between sources.
The classification process includes:
- A faceted taxonomy with 16 categories and about 400 terms, covering product areas, features, workflows, objects, user roles, lifecycle stages, compliance, issues, and technical components.
- Rules for required tags, so every source was findable by product area, workflow, and purpose, and troubleshooting and integration sources carried the detail they needed.
For support tickets specifically: instead of treating tickets as troubleshooting-only material, the skill surfaces how-to steps, warnings, best practices, and conceptual explanations that would otherwise stay buried in support conversations. Because tickets are high volume, a Python script extracts the content so the company isn't paying for tokens to do it.
Why I built it
It's expensive for a drafting skill to find and read up to hundreds of files on every run. I needed a way to ingest each piece of content one time only, classify it, and index it so drafting runs know exactly where to find the information they need and read only the information they need.