# How Daniel Rusnok flipped his Substack reel pipeline to ask authors before rendering

Daniel Rusnok runs Reads in Motion, which turns other writers' Substack essays into short vertical videos on one RTX 3060 in his home office. Under the old order, he rendered each reel first and then asked the author for permission to publish. Fifty-three reels were rendered and 9 were published. He reversed the order. The pipeline now does a cheap close read and builds a storyboard with real render frames and a voice sample. An automated message sends that storyboard to the author, and the expensive render starts only after the author says yes.

| | |
|---|---|
| **Who** | Daniel Rusnok, software engineer and solo builder behind Digital Craft Workshop and Reads in Motion. Substack: @danielrusnok (verified 2026-10-17: author metadata on the post; byline reads "Daniel (Dig. Craft Workshop)") |
| **Problem** | Most of his GPU time went to reels that authors never approved, because the permission ask came after the render |
| **Stack** | One RTX 3060, a nightly render job, Claude's largest model (close read and script), ElevenLabs (paid voice service), a storyboard review page, Substack DMs, a submission form |
| **Confidence** | Medium (first-person; all figures predate the change, and no post-flip results are stated) |
| **Last verified** | 2026-10-17 |

## The problem
Every night a job picked twelve posts and ran each one through five stages. The stages were: write a script, generate a still per scene, animate the stills, add voiceover with captions, and compose the finished 30-second video. One reel takes roughly 47 minutes. He writes, "There is one graphics card and it does one job at a time."

Only after the render did he message the author to ask whether he could publish. Of the last 55 authors he asked, 21 replied and 10 said yes. Nobody said no. He explains the silence this way: "Substack DMs from a stranger sit in a request folder that most writers never open." The card spent nine hours a night on twelve reels, and forty-three reels "rendered for nobody."

He says the finished video was hard to give up as the pitch, and authors who replied "were usually delighted by it." His conclusion: "But the arithmetic beats the delight."

## The blueprint (as he describes it)
1. **Do the cheap part first.** The pipeline reads the essay, writes the script and generates the still image for each scene. He says this takes "a few minutes and almost no time on the graphics card."
2. **Close-read the essay with a large model.** Before any scene is written, Claude's largest model reads the whole essay and works out four things:
   - the thesis
   - the emotional arc
   - the strongest verbatim lines
   - "the one passage that makes a reader forward it to a friend"
3. **Cast a voice per essay.** He replaced the free local text-to-speech model with ElevenLabs. The voice is now cast from the close read instead of one fixed voice for everything. His examples are "A wry British delivery for a comedian's memoir, something deep and level for a finance argument." The close read also sets how expressive the read is and how fast it moves.
4. **Write the script from the same read.** Scripts used to come from "a small, cheap model" and he says "They summarised." Now the script follows three rules:
   - Verbatim quotes stay verbatim.
   - The hook comes from the strongest moment in the piece, not the opening paragraph.
   - Hooks are written as open loops. He took this rule from his history channel, where "an open loop beats a shock."

   The on-screen card must add a second layer of information instead of repeating the spoken line, "because most people watch without sound."
5. **Build a storyboard review page.** The stills are composed into a page, one scene at a time. The frames come from the same rendering code the final video uses, with the same background, header and on-screen card. Next to each frame are three items:
   - the spoken line
   - the on-screen text and which kind of card carries it
   - the camera move the scene gets when it animates

   A play button at the top reads the author's actual opening line in the cast voice. He says the page takes about a minute to read. In his words: "If the angle is wrong, it is wrong before the machine has burned an hour on it."
6. **Ask first, openly automated.** The message to the author is a short note about the piece with a link to the storyboard and a question. Because the close read now happens before the ask, the note can point at a specific line instead of opening with a generic compliment. The message says it is automated in its opening tag and first line, and he reads every reply himself. He has not raised the message volume to whatever Substack tolerates, and he does not plan to. His reason: "A request folder full of automated messages is the exact problem I am complaining about here." The cap is now set by how many writers are worth asking.
7. **Render only on a yes.** If the author says yes, the render starts that hour. The finished reel lands in the same conversation, and he does not publish until the author has seen it. "Say nothing and nothing happens."
8. **Offer an inbound route.** Writers can submit a post through a form. Submissions skip the queue, and the same storyboard-first rule applies.

## Stated results
- Under the old render-first order: 53 reels rendered and 9 published. One more is waiting on a slot, and forty-three rendered for nobody.
- Of the last 55 authors asked after rendering, 21 replied and 10 said yes. Nobody said no. Thirty-four never saw the message or moved on.
- From his separate history channel: an open-loop hook "held viewers more than twice as long" as one that simply stated something dramatic, with the same production and length.
- After the flip: None stated. The post says the new order starts "today."

## Caveats
- He sells a Reel Pipeline Kit on Gumroad, which the post links to. Reads in Motion is his own project, and the post promotes its submission form.
- The post was published 2026-07-19 and describes a change starting that day. It contains no data on reply or approval rates under the new ask-first order.
- The 55 authors are the last 55 he asked, not the full set behind the 53 rendered reels. The TL;DR describes the outcome as reaching "one reader in six." That framing is his.
- He does not name the specific Claude model or the image and animation models. He does not state the ElevenLabs cost.
- The history-channel hook finding comes with no sample size or method.
- The byline shows only "Daniel." The full name comes from his Substack subdomain and handle.
- Handles: Substack @danielrusnok (verified 2026-10-17: author metadata on the post; byline reads "Daniel (Dig. Craft Workshop)"). LinkedIn: a matching profile turned up in search but is not linked from the post, so it is not used. X: not checked, no handle used.

## Sources
- https://danielrusnok.substack.com/p/53-reels-rendered-9-published-so (primary: first-person build log on his Substack)

---
*Want agents like these working your distribution? [See the Agent Stack Teardown](/agent-stack-teardown). Get new teardowns weekly: [subscribe](/subscribe) (double opt-in).*