Video Digest

Turn any public video URL into a transcript-backed digest: summary, key points, action items, and three follow-up questions personalized to the user.

Requirements

• SUPADATA_API_KEY environment variable (from https://dash.supadata.ai/organizations/api-key). If it's missing, the script exits with a clear message - relay it to the user and stop; don't guess at the video's content.
• Supported sources: YouTube, TikTok, Instagram, X (Twitter), Facebook, and public file URLs (MP4, MP3, WAV, etc., up to 1 GB). Only publicly accessible videos work - private, paywalled, age-restricted, or login-required videos return 403/404.

Workflow

1. Fetch the transcript

Run the bundled script (handles auth, async job polling for videos over ~20 minutes, and error mapping):

python scripts/get_transcript.py "<VIDEO_URL>" --out transcript.txt

Optional flags: --lang en (preferred transcript language), --mode native (existing captions only - cheaper, no AI generation) or --mode generate (force AI transcription). Default auto tries native first.

Read transcript.txt afterward. For very long transcripts (>50k chars), read in chunks and build the summary incrementally. Long videos process asynchronously - the script polls and this can take a minute or two; that's normal.

If the script errors, report the plain-language cause (missing key, restricted video, no transcript available) and suggest the fix. Don't summarize a video you couldn't transcribe.

2. Write the digest

Deliver in chat, in this shape:

## <Video title or topic, inferred from transcript>

**Summary** - 2-4 paragraphs capturing the core argument or narrative, in the
speaker's framing. Prose, not bullets.

**Key points** - 4-7 bullets, each a substantive claim or insight from the
video (not chapter titles). Include specific numbers, names, or examples the
speaker used.

**Action items** - concrete things a viewer could do based on the video's
content. If the video isn't instructional, extract implications instead
("worth watching X", "reconsider Y"). Skip filler - 2-5 items.

Ground everything in the transcript. Don't pad with knowledge the video didn't cover, and don't soften or reinterpret claims the speaker made - the user wants to know what this video says.

3. Ask 3 personalized questions

End with three questions connecting the video to the user. Personalization source, in order:

1. Memory files. Check what's available in this session: a memory directory or memory tool, CLAUDE.md / user context files, or files like MEMORY.md, memories/*.md in the working or home directory. Read whatever exists and look for the user's projects, role, goals, and interests.
2. Conversation context. Anything the user has said this session about their work or aims.
3. Fallback. If neither yields anything, ask three natural next questions about the video's topic - the questions a curious viewer would ask after watching (deeper dives, counterpoints, practical application).

Good personalized questions bridge the video to the user's actual situation: "The speaker recommends X for early-stage teams - would that fit the pipeline you're building at <their project>?" Bad ones just restate the video or bolt the user's name onto a generic question. If memory gives you only partial context, blend: one or two personalized, the rest topic-driven.

Present the questions as a short numbered list under a "Your turn" heading, and invite the user to answer any of them or ask their own.

Notes

• One native transcript = 1 credit; AI-generated = 2 credits/minute. Prefer auto mode unless the user asks otherwise.
• If the user gives multiple URLs, digest each separately in one response, with one combined "Your turn" section at the end.
• If the user only asks for the raw transcript, run step 1, share the transcript file, and skip the digest - but offer it.