Transcript Intelligence

The average knowledge worker spends 4+ hours a week in meetings and captures almost nothing actionable from them. Meeting notes, when they exist, are scattered across Drive folders with inconsistent structure, no tagging, and no follow-through. Insights stay trapped in transcripts nobody re-reads. This project builds a fully automated pipeline that converts raw meeting transcripts into structured intelligence documents, without any manual step between recording and output.

1. 1.
   Automated Trigger: Watch a Google Drive folder for new transcript files and trigger the analysis pipeline automatically on each new upload, no human initiation required.
2. 2.
   Structured Output Schema: Design a document schema that captures what teams actually need to act on: key insights, decisions made, pain points raised, and next steps with owners.
3. 3.
   LLM Constraint Design: Constrain Claude to produce the defined output schema reliably, not allow it to summarise in whatever format felt natural to the model.
4. 4.
   Drive Integration: Write the structured output back to Drive as a formatted Google Doc, in the same folder structure as the source transcript, with consistent naming.

• ◆
  Schema Design First: Before writing any prompt engineering, defined the output schema by asking: what does a team need from a meeting summary to actually act on it within 24 hours?
• ◆
  Drive API Integration: Implemented a Google Drive folder watcher using the Drive Changes API, with file type filtering to avoid triggering on non-transcript uploads.
• ◆
  Claude Prompt Engineering: Engineered a structured output prompt that forces JSON-compatible section extraction, tested across 40 historical transcripts before production deployment.
• ◆
  Output Formatting: Used the Google Docs API to write structured summaries with heading hierarchy, bold key terms, and consistent section ordering matching the schema.

The natural instinct is to prompt the LLM to 'summarise this meeting' and let it decide what to include. That produces summaries that feel thorough but are nearly impossible to act on, because each summary is structured differently and prioritises different things. The schema-first approach means every output answers the same questions in the same order: What was decided? What pain points came up? Who owns what next step? Teams can scan it in 90 seconds.

Meeting transcripts vary widely in quality, some are clean Zoom auto-transcripts, others are noisy Otter.ai files with speaker attribution errors and filler words. The preprocessing step normalises these before passing to Claude: strip filler patterns, fix speaker attribution where possible, and chunk long transcripts to stay within context limits while preserving logical meeting segments.

I designed the output schema before touching the Claude prompt. What does the team actually need from a meeting summary to act on it? That question drove the structure. The LLM was constrained to produce that format, not left to generate whatever felt natural.

Eliminated the manual meeting-notes process entirely. Structured intelligence is available in Drive within minutes of a meeting ending.