creator-intake 开源：把 YouTube 频道变成可验证引用的问答页面

Point it at a YouTube channel and it builds a page where people type what's going on in their own words and get:

1. A straight answer, grounded only in what the creator has said on video.
2. The clip where they said it, playing from that exact second. Every quote is checked word for word against the creator's own captions before it shows. Interviewers, guests and AI voices are screened out, so it never puts words in the creator's mouth.
3. The creator's next step: a call, a product, a course. A decision model picks the one that fits what this visitor asked and did, and asks them one quick question when it can't tell, instead of guessing.

It answers about what the creator said and shows the receipt. It never talks as them.

• Live example: https://www.connorgallic.com/ask
• How it works, with real numbers: https://www.connorgallic.com/ask/for-creators
• Build a preview from any channel: https://www.connorgallic.com/ask/new

About $0.002 a question with GPT-6 Luna on OpenRouter, and about $0.00004 per decision.

How it works

 visitor types ──► Jev reads it (topic, intent, which cards)      0.2 s ┐
               └─► the writer drafts the answer from every clip   ~7 s  ├─► page built for them
                    └─► code checks each quote word for word             │
                        └─► Jev checks each clip is on point      0.3 s ┘
 visitor taps / opens a clip ──► Jev re-reads, page reshapes (no new answer)

• The writer (any OpenRouter model, GPT-6 Luna by default) gets every caption line of the creator's videos in one cached prompt and drafts the answer, quoting clips by id.
• api/_verify.js keeps a quote only if those exact words appear in one unbroken run of the creator's own caption lines. That run also gives the second to play from.
• Jev (TypeSafe's "System One" decision model, served by OpenRouter at /api/v1/systemone) never writes text. It answers typed questions with calibrated odds. The page acts on a topic or intent only when Jev is at least 60% sure and at least 20 points ahead of its second choice. Otherwise it asks the visitor. What it infers changes the layout, never the wording of the answer.
• Custom offers: give a creator several offers (a course, a call, a product), each with a one-line when. Jev scores every offer against what the visitor asked, and the clear winner becomes the button. Marc Lou's preview picks between ShipFast, DataFast, CodeFast and TrustMRR this way. Set them with build/set-offers.mjs.
• The page is a static HTML file that renders a component list from the API (A2UI-style). "Why this page?" on every answer shows what was decided and how sure it was.

Self-serve: paste a channel, get a page

 /new  ──POST /api/new──►  job (Vercel Blob)  ◄──poll──  build/worker.mjs (any box with yt-dlp)
                                                           1. read the channel
                                                           2. captions for the newest 60 uploads
                                                           3. keep only the creator: a vision check of each thumbnail
                                                              against the avatar, AI-voice phrases, re-uploads
                                                           4. draft the page (intro, topics, starter questions)
                                                           5. Jev tags each clip's topic
 /c/<slug>  ◄──────────────────  creators/<slug>/*.json  ◄──  6. publish

The worker never holds the OpenRouter key. Every AI call goes through the app's narrow proxy (api/ai.js), which is locked to a shared WORKER_SECRET, to two endpoints, and to an allowlist of models.

Anyone can paste anyone's channel, so a self-serve page is an unofficial preview. It's labelled as one, it's kept out of search engines, and its only next step is the creator's own channel until the owner claims the page.

Run your own

1. Deploy this repo to Vercel. In the dashboard, add a Blob store to the project: Storage → Blob, which sets BLOB_READ_WRITE_TOKEN.
2. Set the env vars:

1. Run the worker anywhere that has Node 20 and yt-dlp:

APP_URL=https://your-app.vercel.app WORKER_SECRET=... node build/worker.mjs

Or build one page by hand, with full control

cp -r creators/_template creators/<slug>          # edit config.json: offer, cards, topics, routes
node build/fetch.mjs <slug>                        # captions (yt-dlp)
node build/index.mjs <slug>                        # corpus.json + featured.json
python build/contact_sheets.py creators/<slug>/unscreened.json /tmp/sheets   # screen who's on camera
OPENROUTER_API_KEY=... node build/tag-lanes.mjs <slug>   # Jev tags each clip's topic
node build/stats.mjs <slug>
node build/use.mjs <slug>                          # serve it from data/
npm test && vercel deploy --prod

creators/connor/ is a complete worked example: 548 screened clips, cards, topics, routes and demo snapshots. creators/_template/config.json documents every field.

To serve the page under another site's path, add a proxy rewrite, for example in Next.js: { source: "/ask/:path*", destination: "https://<app>.vercel.app/:path*" }. The page detects the sub-path by itself.

Files

Tests

npm test

These cover the quote guarantee, the spend cap, Jev's thresholds, routing and fallbacks, and channel parsing.

License

MIT