bertbertov Transcript Analyzer
Overview
Analyze meeting / podcast / interview / YouTube transcripts using Cerebras llama-3.3-70b to automatically extract and categorize:
- Decisions — explicit agreements or choices made
- Action Items — tasks assigned to people
- Opinions — viewpoints expressed but not agreed upon
- Questions — unresolved questions raised
- Terms — domain-specific terminology for incremental glossary
Prerequisites
1. Install dependencies (one-time)
cd ~/.claude/skills/transcript-analyzer/scripts && npm install
2. Get a Cerebras API key
Sign up at https://cloud.cerebras.ai — there is a free tier with generous limits (this is what makes llama-3.3-70b extraction effectively free for personal use). Create a key in the dashboard, then put it in scripts/.env:
CEREBRAS_API_KEY=csk-...
Reference doc: https://inference-docs.cerebras.ai/
Usage
cd ~/.claude/skills/transcript-analyzer/scripts && npm run cli -- <transcript-file> -o <output.md> [options]
Options
| Option | Description |
|---|---|
<file> |
Transcript file to analyze (first positional arg) |
-o, --output <path> |
Write markdown to file instead of stdout |
--include-transcript |
Include full transcript in output [default: off] |
--no-extractions |
Exclude extractions section |
--no-glossary |
Exclude glossary section |
--glossary <path> |
Custom glossary JSON path |
--skip-glossary |
Don't preload glossary terms |
--max-terms <num> |
Limit glossary suggestions |
--chunk-size <num> |
Override chunk size (default: 3000) |
Examples
Basic analysis
cd ~/.claude/skills/transcript-analyzer/scripts && npm run cli -- /path/to/meeting.md -o /path/to/analysis.md
Include original transcript
cd ~/.claude/skills/transcript-analyzer/scripts && npm run cli -- /path/to/meeting.md -o /path/to/analysis.md --include-transcript
Extractions only (no glossary)
cd ~/.claude/skills/transcript-analyzer/scripts && npm run cli -- /path/to/meeting.md -o /path/to/analysis.md --no-glossary
Analyze a specific section
sed -n '50,100p' /path/to/meeting.md > /tmp/section.md
cd ~/.claude/skills/transcript-analyzer/scripts && npm run cli -- /tmp/section.md -o /path/to/section-analysis.md
the user-specific assets
Existing transcript corpora on this machine (point the CLI at any of these or your own paths):
- 451 DaytradeWarrior YouTube transcripts — bulk-downloaded corpus, ~2M words. Use to extract Warrior's recurring trade rules, setups, risk decisions, and build a strategy glossary. Path lives under
C:\Users\A\Desktop\YouTube\...(point at the actual transcript files when invoking). - MIF book knowledge bases — 541 books distilled at
C:\Users\A\Desktop\МИФ\_knowledge_bases\. Useful for opinion / decision extraction across long-form content. - Lagarra meeting recordings — when transcribed, route through this skill before pushing summaries to assistants.
When running against the DaytradeWarrior corpus, prefer --no-glossary per-file then merge glossaries at the end (the default data/glossary.json is incremental).
Output Format
The tool generates markdown with:
- YAML frontmatter — processing metadata: chunks processed, extraction counts by type, new terms discovered, model used (llama-3.3-70b via Cerebras), token usage (input / output / total).
- Extractions — categorized findings with confidence scores; each includes speaker (if identified), source snippet, and related terms.
- Glossary — approved terms from the existing glossary plus suggested new terms with definitions.
Files in this skill
scripts/cli.ts— main CLI entry pointscripts/src/lib/extract-service.ts— AI processing logic using Cerebrasscripts/src/lib/markdown.ts— markdown output generationscripts/src/lib/term-utils.ts— term deduplication utilitiesscripts/src/lib/mockExtractor.ts— mock mode for offline testingscripts/src/lib/utils.ts— shared helpersscripts/src/types/index.ts— TypeScript type definitionsscripts/data/glossary.json— default incremental glossary storescripts/package.json,scripts/tsconfig.json— Node/TS config
Source
Adapted from https://github.com/glebis/claude-skills/tree/main/transcript-analyzer (no license declared — personal use only, do not redistribute).