Auditing a court-rules reference library with parallel AI link checking
One staff member had built a 295-file court-rules reference library by hand over years, and nobody could say how much of it was still accurate. Checking is exactly what a large, informally maintained library resists, so the audit split 272 URLs across ten parallel agents just to get through it, and the real finding wasn't a broken link. It was that nothing could be verified automatically until each jurisdiction had a documented source of truth.
Category
Reference-library audit with parallel research.
What I was trying to accomplish
An honest structural audit of a large, entirely hand-maintained court-rules reference library nobody could vouch for the freshness of, and a plan to get AI checking it for changes.
The data challenges
Walking roughly 295 files across 16 counties, extracting and checking 272 unique URLs when direct web requests were blocked in the working environment, splitting that check across ten parallel agents, and untangling an informal folder-naming convention that was quietly carrying two different meanings in one prefix.
The result
A structural audit of what's solid, broken, duplicated, or missing, a per-jurisdiction source list, and a four-phase roadmap, built on one key finding: automated freshness checking can't work until each jurisdiction has a documented source of truth, so that index has to come first.
Technical details
Tools: Claude, filesystem connector, PyMuPDF, parallel subagents
Over years, one staff member had built a reference library of court filing rules and contacts: about 295 files, built out for 16 Washington counties with partial coverage elsewhere. It was valuable and entirely hand-maintained, and nobody knew how much of it was stale.
Working from a snapshot copy of the library, Claude walked the full directory tree and file metadata through a filesystem connector. In a follow-up session it extracted every hyperlink from the 20 per-county quick-reference PDFs and checked the 272 unique URLs, split across ten parallel subagents since direct web requests were blocked in the working environment. Claude also worked out the library's informal folder-naming convention, and when I explained that the prefixes marked both progress and importance, it pointed out that one prefix was carrying two independent variables and recommended separate priority and status fields.
The outputs were the structural audit, a per-jurisdiction list of source URLs, and a four-phase roadmap: stabilize, index, standardize, expand. Along the way the file connector hung mid-session. After a restart, Claude corrected its own first diagnosis, narrowed the fault to the file-transfer bridge rather than the connector itself, and recommended direct upload as a workaround.
Pull up a chair.
Let's talk through where things stand, and where they could go.