What problem does it solve?
Large datasets like meeting recording archives, transcript collections, and document dumps are too big to load into an assistant's memory directly, yet users still need them searchable. This Skill ingests such corpora by cold-storing raw files in a workspace imports directory and writing only a small set of map pages into memory, keeping the corpus searchable on demand without bloating memory.
Core Features & Use Cases
- Cold-store with credential screening: Copies raw files into
imports/<source>/ after scanning filenames and contents for secrets, excluding anything sensitive before it lands.
- Inventory and slice planning: Runs
scripts/inventory.ts to census the corpus (file counts, sizes, extensions, date ranges) and propose date-windowed slices sized for one skim pass each.
- Bounded map generation: Skims each slice into 10-50 map pages following a strict template, ingests them via
assistant memory ingest with dry-run validation, and authors a drill-in retrieval skill with runnable search scripts.
- Use Case: A user exports a year of Fathom meeting recordings. The Skill lands the transcripts in the cold store, builds quarterly map pages capturing decisions and open threads, and creates a
fathom-lookup skill so questions like "what did we decide in the March planning call" route to the right files.
Quick Start
Ask the assistant to import your Fathom recordings export folder into memory so it stays searchable without pasting everything into chat.
