Heads up — this one is context-hungry. Claude reads and reasons over every measure that isn't auto-templated, so a large model can use a lot of context in a single run. Trim the model first (see the tip above) and run it when you're not deep into a long session, so there's plenty of room to work. The skill processes measures in batches and resumes automatically if a run is interrupted, so big models still finish.

Type /write-descriptions, point it to your PBIP project, and the skill handles everything:

1. Scans all TMDL table files and extracts measure names, DAX expressions, format strings, and existing descriptions

2. Auto-templates deterministic infrastructure measures (hex color constants, blank placeholders, tooltip and legend helpers, standard aggregations) without Claude generation

3. Discovers prefix clusters — groups of measures sharing a common prefix — and processes one representative per cluster for efficient token use

4. Generates business-friendly descriptions for all remaining measures in batches, resuming automatically from any interruption

5. Writes /// description triple-slash comments directly before each measure declaration in the TMDL files

6. Produces a verification report listing every measure, its DAX, and the generated description for human review