Code With Claude

Code w/ Claude — San Francisco 2026 · Full Transcripts

Anthropic uploaded the complete session recordings from their **Code w/ Claude** event in San Francisco — ~8 hours across **19 videos**, free on YouTube. This folder holds a clean, readable transcript of **all 19** talks, each with an AI-generated thesis + key-points summary at the top.

Talks span keynotes, workshops, and demos, featuring Dario & Daniela Amodei, Boris Cherny (creator of Claude Code), Guillermo Rauch (Vercel), Jarred Sumner (Bun), and teams from Cognition, Gamma, Harvey, Datadog, Asana, Google Cloud, Replit, and Cursor.

How these were made

Each transcript was produced from the YouTube audio with:

• Deepgram Nova-3 — pre-recorded transcription with smart formatting, punctuation, and paragraphing, automatic language detection, no chunking (handles long audio natively). Blazing fast — the opening keynote's 47 minutes transcribed in ~5 seconds (530× realtime). This is the base text: its wording, punctuation and paragraphing are what you read.
• ElevenLabs Scribe v2 — a second transcription of the same audio, with keyterm prompting for this event's vocabulary. Used only to correct the base text, never to replace it: the two are aligned token by token and Scribe's reading is adopted only where it supplies a known-correct domain term or fixes a known corruption. Also supplies speaker diarization and the word-level timings behind the timestamps and chapters.
• Gemini Flash-Lite — generates the ## Thesis + ## Key Points summary block at the top of each file.

**Why the second pass.** Nova-3 is excellent on ordinary speech but does not know the product names this event is about. Across all 19 talks it corrupted 22.96% of domain terms (163 of 710) — Claude's own name appears as *Cloud*, *quad*, *Clot* and *clawed*, and CodeRabbit is spelled three different wrong ways. Scribe v2 with a 116-term keyterm list has an observed corrupt