StrandScribe: Turning Short Videos Into Articles Where Every Claim Has a Source
Share a reel, see what each claim rests on, edit, post. Claim checking against fetched pages, a durable Cloudflare pipeline, a billing ledger, and a threat model.
- YEAR
- CATEGORY
- SaaS AI security Vue Cloudflare Nuxt automation LLM
- BY
- Aleksei Skorjak
- FILE
- strandscribe-video-to-sourced-article.prj
StrandScribe turns a short video into an article in which every factual sentence points at a page that was fetched and read.
A reel can build an audience that a search engine cannot read
Short video is where many creators build their following, and none of it is visible to Google or to an AI assistant answering a question. A written version would be. The quick way to get one is to paste a transcript into a chatbot, which returns fluent text and no promise that any sentence in it is true. The creator then publishes it under their own name.
StrandScribe is built for that gap. You share a reel, it lists the claims the video makes, finds pages that support or dispute each one, and lets you choose what the piece is for before it writes a word. You edit the draft and post it to your own blog. The product is live at strandscribe.com, with three reels free and no card, then Starter at 9 euros a month for 50 reels or Pro at 29 euros for 250.
How a claim becomes a sentence
The audio is transcribed and reduced to a short list of checkable claims. Opinions, jokes, and calls to follow the account are dropped. Each claim carries a "heard clearly" score, so a name or number the speech model may have misheard is visible before anyone searches for it.
Search comes next, and it only produces addresses. Nothing a search engine says about a page counts as evidence. The application fetches and reads every candidate page, and a source is kept only when a passage bearing on the claim appears word for word in the fetched text. If the passage cannot be found, the source is dropped. Pages that dispute a claim are kept as well, with the exact passage, because a person about to repeat a claim should see the page that contradicts it.
Each kept source gets a trust score based on the publisher and on how many independent sites say the same thing. In the example on the site, an engineering blog scores 60 and a business marketing page from the same company scores 30 for the same sentence. The score is a guide to where to look first, and the product does not present it as proof that a page is right.
The article writer receives only the kept passages. It states as fact what they say, numbers each of those sentences by its source, and leaves unsupported claims out. The source list is assembled from the kept passages, so the model never writes a citation.
The connecting prose between numbered sentences is still written by a language model, and that is where an unsupported detail can slip in. The product says so and tells the user to read the unnumbered sentences first.
The model proposes, the code verifies, the owner publishes
That sentence is the first invariant in the architecture decision record I wrote for the AI pipeline. The others follow from it. Search results are discovery and never evidence. The application retrieves the canonical page itself and binds each accepted span to the page's content hash. A model's confidence in its own draft is never enough, so ambiguous cases go to deterministic checks, an independently configured validator, or a person.
I also declined to pick a provider on instinct. The record specifies an acceptance harness: 80 adjudicated cases split 60 and 20 into development and holdout sets, candidate pipelines compared against each other, and a timed human baseline of someone doing the same job by hand with a free chatbot. The first live pilot has a spending ceiling of 5 US dollars. Staging uses Tavily for discovery and Groq for speech and extraction, and the record marks those assignments provisional until measurements say otherwise. The quality gates stay fixed and the models can be swapped.
Three trust zones on Cloudflare Workers
StrandScribe runs on Cloudflare Workers with Nuxt 4, Vue 3, and strict TypeScript. I designed the runtime around three trust zones instead of a fleet of services: the web application, a webhook ingress on its own hostname that accepts Instagram and payment events on exact routes only, and a non-public pipeline for extraction, research, drafting, and publishing. The split keeps the one public exception, events arriving from outside, away from everything else.
State lives in D1, and large private artifacts live in R2. Any change that starts asynchronous work writes the aggregate, the operation, and an outbox row in one atomic step. Queues absorb at-least-once delivery with dead-letter queues, and consumers deduplicate on stable operation IDs. Cloudflare Workflows own one durable operation each, and the repository has separate workflows for source briefs, research, article drafting, automation, and diagnostics. No job depends on waitUntil() as its only protection against loss. Forty-two migrations took the schema from sessions and webhook receipts to scheduled posts, voice profiles, and invite lists.
Money, access, and what happens on a replay
Payments run through Creem. Entitlements come only from verified provider events and never from a browser redirect, so closing the tab after paying changes nothing and a forged redirect grants nothing. Underneath the plans sits an append-only ledger where corrections are compensating entries. Publishing reserves before the attempt, settles only on confirmed success at the destination, stays pending reconciliation when delivery is ambiguous, and releases exactly once after a confirmed failure.
Webhook handling follows the same discipline. The signature is checked on the raw body with a constant-time comparison, body size is capped, and a durable unique receipt means a replayed event is recognized instead of processed twice.
Sign-in is Google only. State, PKCE, and nonce are checked, and the session is one opaque 256-bit cookie that scripts cannot read, stored hashed. Keys for connected publishing platforms are verified with the platform, encrypted with AES-256-GCM under a key held apart from the database, and never displayed again, even to their owner. The Instagram sender identity is stored as a one-way hash.
A product that reads the open web for a living has its fetcher as a major attack surface. Pages are fetched only from public https addresses, and private ranges and redirects into them are refused. Text on a fetched page is treated as material to quote and never as instructions, which is the defense against a hostile page telling the model what to write. Shared videos sit in private storage, are read once, and are deleted, with 24 hours as the backstop.
Getting a reel in without scraping Instagram
Users can paste a link, upload a file, or share a reel to the StrandScribe bot on Instagram or Telegram. The Instagram path uses the Messaging API and receives only what Instagram delivers for that one message. There is no crawler and no bulk export of anyone's account. Linking a sender to an account takes a single-use, short-lived code and an explicit confirmation inside the signed-in app, so nobody can attach a stranger's messages to their own account by guessing codes.
The how-it-works page was written for two readers: a creator deciding whether to trust a DM, and a platform reviewer deciding whether to trust the Messaging API. Both get the same walkthrough.
One video, several pieces, in your own voice
The same verified facts can become a news article, a feature, an opinion piece, a newsletter issue, or a thread. StrandScribe proposes angles from the claims, and the user decides what the piece is for.
Voice is stored as traits: sentence length, openings, person, humor, structure, and vocabulary to prefer or avoid. They come from up to five pages of the user's own blog, and no sample text is kept or quoted. The same brief written in the house voice and in a first-person test voice averaged 20 words per sentence against 9, which gave me a measurable check instead of a feeling.
Publishing reaches about 30 destinations, including Ghost, WordPress, Beehiiv, Notion, Bluesky, Mastodon, Telegram, and LinkedIn. In normal use nothing posts until the user presses the button. Autopilot is the opt-in exception, with off, draft, and publish modes, and it stops in the inbox with a stated reason when no supporting source reaches medium trust, a stage fails, or a platform is disconnected.
The business around the code
A product also has to be priced, billed, supported, and explained to the platforms it depends on. Pricing is public and flat, quoted in euros or US dollars, with tax added at checkout where it applies and cancellation from the account. The site carries a privacy policy, terms, an acceptable use policy, a data deletion page, a refund policy, and a security page with a disclosure address. The repository includes a waitlist and an invite system, so access can be opened in controlled steps.
The company blog is the same bet the product sells: written pages that search engines and assistants can read. Eight articles answer the questions people ask before they need a tool like this, including how to fact-check AI-written content, whether AI content hurts search ranking, how to get cited by ChatGPT, and how to turn an Instagram reel into a blog post.
How the work was run
The git history runs from 24 August to 13 October 2026 across 66 commits and roughly 58,000 lines of TypeScript, Vue, and scripts, tests included. Every code directory has a DIRECTORY.md stating why it exists, who owns it, its boundaries, its security and data considerations, and the command that verifies it, and a lint script checks that each one is present.
Development used AI coding agents under a written execution policy: one independently verifiable outcome per task card, named allowed paths, acceptance checks, explicit non-goals, and at most one fallback attempt before reporting a blocker. The agents wrote code, and the boundaries they worked inside were mine.
The tests cover what breaks money and trust first. There are more than 400 test cases across authentication, the billing ledger and its webhooks, Meta ingress, research, drafting, publishing, and voice, plus browser tests with axe accessibility scans and visual snapshots.
What StrandScribe says about how I build
StrandScribe is a complete product and not a demo: an AI pipeline whose output can be audited, payments that cannot grant or charge twice, authentication and webhook handling written against a threat model, a fetcher that treats the open web as hostile, a compliance walkthrough for a platform reviewer, public pricing, legal pages, and a content plan.
The common thread is deciding where a system is allowed to trust something. Models propose and code verifies. Verified events grant entitlements and redirects do not. A fetched page is material to quote and never an instruction. I write those boundaries down before the code and then test them.
Live product: strandscribe.com