Auto-load meeting transcripts into Claude Code on session start (SessionStart hook)

A Claude Code SessionStart hook fires when you launch CC. The hook script reads Granola's local transcript store, pulls the day's calls, and injects them into the session along with a classification prompt that tags each transcript by deal / company / topic. Result: every CC session opens already aware of recent meetings — no copy-paste, no "let me catch you up." Mac-only as currently shipped.

Founders running 4+ calls per day repeatedly hit the same friction: every CC session starts cold, every follow-up requires re-pasting context, every CRM update needs the operator to summarize. The hook removes that step entirely; the agent already has the morning's meetings classified before you type the first prompt.

1. Write extract-granola.ts — script that reads Granola's local transcript files from disk and emits the day's transcripts as text (Granola stores them in user library on Mac).
2. Write load-context.ts — the SessionStart hook entry point that calls extract-granola.ts, wraps the output with a classification prompt (assign each transcript to deals/companies/topics), and prints to stdout for CC to ingest.
3. Register the hook in ~/.claude/settings.json under the SessionStart event, pointing at load-context.ts.
4. Adapt the classification prompt to your taxonomy (your deal pipeline, your team, your projects). The default is opinionated and won't match yours.
5. Test: open a new CC session and verify the boot output shows the day's calls grouped by category.

The classification prompt is opinionated to one operator's deal/company taxonomy and will produce noise out of the box for anyone else. Strip the classification section first, confirm the raw transcript injection works, then layer your own taxonomy back on top. Also: if you have 10+ calls/day the injected context can blow past the session-start budget — cap by recency or by attendee.