Blog to video in Make: a Claude agent turns each RSS post into a voiced vertical short (free scenario template)

Hey Make Community, turning a blog post into a short video takes a lot of hand work: a script, a voiceover, scenes, timing and captions for every post, so most posts never get a short at all. So I built a free Make scenario that turns posts from any RSS feed into a 32-second vertical short with a voiceover, captions and music, ready for YouTube Shorts, Instagram Reels or TikTok. A Claude agent reads the post, writes the script and designs every scene, and Shotstack renders it and saves the MP4 to your Google Drive. You run it, and about half an hour later the short is in your folder, named after the post.

The blueprint, ready to import:

AI video Shorts from an RSS feed - Claude agent.blueprint.json (38.0 KB)

How the scenario works

It’s one scenario, and it runs in this order:

  1. You put your blog’s feed address in the RSS step and click Run once. Shotstack’s blog feed is filled in as the example.

  2. The RSS module reads the newest post, and an HTTP step fetches the post’s page, because most feeds only carry a short summary.

  3. A Claude agent (Claude Opus 5.5, through Anthropic’s Managed Agents) gets the post and a brief in plain English. It works in its own cloud sandbox with Shotstack’s command-line tool and its agent skill installed. It writes the script first, then builds the video as one Shotstack edit. It screenshots every scene in a browser and fixes what looks wrong, and checks the edit with Shotstack’s validator until it passes.

  4. Make checks on Claude every 30 seconds, then downloads the finished edit.

  5. Shotstack’s Render from JSON module renders it. Shotstack generates the voiceover, the music and the photos inside that one render.

  6. Make checks the render every 30 seconds, downloads the MP4 and uploads it to your Drive folder. If Shotstack says the render failed, a router writes <post title> - failed.txt to the same folder with the reason instead.

Claude’s sandbox has no Shotstack key, so Claude can only check the edit. The render, and what it costs, stays in Make on your own Shotstack connection.

The first run also sets itself up: if your Anthropic account doesn’t have the agent yet, one route creates it (“Shotstack video editor”) and its sandbox, and every run after that reuses them.

Here’s the full brief Claude gets for each post:

Make a 32-second vertical Short (mp4, 1080 x 1920, 30 fps) that tells the main idea of this article to someone scrolling past. It must look made by hand in a design studio: calm, precise and premium, never "AI art".

Title: (the post's title, from the RSS module)
The feed's summary: (the post's summary, from the RSS module)
The article's web page, as text: (the post's page, from the HTTP step)

The system (copy it exactly)
- One look for the whole video. Light: background #E9E6E1, text #1C1C1C, muted #8C8882. Or dark: background #0D0F13, text #F4F2EE, muted #7D8088. Pick the one that suits the article. Add one accent colour that suits the subject, for example a soft pink #F28DA6. It's only for small things: the spoken word in the captions, the hook's highlighter bar, check badges, the product name in a title bar, a thin line. Never a whole background, except on the end card.
- Three zones, the same on every scene, so nothing collides: the top zone (y 80 to 560) holds the kicker and the headline, the middle zone (y 600 to 1400) holds the scene's visual, and the bottom zone (y 1480 to 1760) holds only the captions. 84 px left and right margins. Few words: besides the kicker, the headline, the captions and the end card's page, at most 8 more on screen at once.
- The top zone: a kicker at y 120, JetBrains Mono, 26 px, uppercase, letter-spacing 0.12em, in the muted colour. On the first scene it's the site's name and the article's date ("SHOTSTACK BLOG · 21 OCT 2026"), then a section number and name ("01 — TEMPLATES"). Under it from y 170, the headline: Inter, weight 600, 104 to 116 px, letter-spacing -0.03em, line height 1.05, sentence case, left-aligned, at most 3 lines and 7 words. It's the biggest, boldest text on screen; only the one huge number is bigger. Its lines rise in one after another, 0.08 s apart.
- 7 or 8 scenes, one idea each, each showing exactly what its sentence of the voiceover says, and timed to that sentence (see Sound and captions). Scenes cut. Nothing flashes, shakes, spins or punches in. Things build up step by step and then hold still long enough to read. A scene's first frame is never empty: its kicker is there from 0 and its headline starts rising at 0.
- Each scene is one html5 clip covering the whole frame, and everything designed for that scene lives inside it: the kicker, the headline, cards, diagrams, numbers and labels. Its background is the look's colour, or transparent where a photo shows through: the end card waits underneath, so a scene without a photo always paints its own background.
- Use these kinds of scenes, in the order the story needs:
  - The opening: the hook. Not the article's title but a 3 to 6 word line ending in a full stop that makes a scroller stop: the article's news, or its sharpest claim or problem ("Gemini 4 Argon just dropped.", "Some data can't leave the building."). It's the biggest headline in the video (Inter 600, 120 px, up to 3 lines), its lines rising in as the voice says them, with an accent highlighter bar drawing in behind its key words. Under it, in the middle zone, the article's subject shown literally, never a symbol or metaphor of the idea (no speech bubbles for languages, no lightbulbs for ideas): what the article makes or runs, recreated in html5 with the article's real names and data, for example the video it outputs playing on a phone, its app's screen, or its config. A photo may fill that phone's or window's screen when it shows something real from the article (its device, its place, the subject of its example): an image clip sized to the screen, which the html5 leaves as a transparent hole (a transparent screen element whose huge spread box-shadow paints the background, so the scene's body stays transparent).
  - Product-style UI that shows how the idea works. A dark app window (rounded 24 px corners, a mono title bar with three dots, the product's name in the accent), with a request typing itself into a bubble. A row of white rounded tiles (200 px, thin #DDD9D3 border, soft shadow), each with a simple line icon, a label (Inter, 40 px) and a mono sub-label (26 px, muted), joined by thin lines that draw themselves, and an accent check badge on each tile as its step is done.
  - Everything in the middle zone must be readable on a phone: no text under 26 px, and the app window's text at least 32 px. Each scene's visual fills most of the middle zone (at least 700 px wide or 500 px tall), never a small thing adrift in empty space.
  - A list that builds one line at a time (a small icon, then Inter at 64 px), with the last line muted.
  - One huge number or word (Inter, 360 to 480 px, inside the middle zone) with a short mono label under it. Use it once, for the article's most striking fact.
  - A real object in a photo, with a small floating label card that names it.
  - The end card: the accent as the background on the light look, or the look's background on the dark look. The headline "Read it on" and the site's name, and under it a recreation of the article's page in a white browser window, built in html5 from the page's real text: the address bar with the article's address, the site's name and menu in its header, then the section and date, the article's title, and its first few contents headings. No photo. It doesn't animate: it's one html5 clip with start 0 and length "end" on the track under the photos. The scenes and photos cover it until the last sentence, which plays over it. Never give "end" to a clip that starts later than 0: Shotstack makes its length the whole timeline's.
- Photos: 1 or 2, each generated with nano-banana-2 (options aspectRatio "9:16", resolution "2K"). Each is a real-looking photograph of something real in the article (its device, its place, the subject of its example), never a symbol of the idea: one subject on a seamless backdrop in the look's colours, or a real place, soft daylight, real shadows, shallow depth of field, shot on a medium-format camera, with the top and bottom zones plain, so the headline and captions stay clear. On the dark look it can be a quiet real place at dusk with one warm light. Never: screens showing UI, people at laptops, holograms, glowing network lines, particles, robots, circuit boards, purple-blue gradients, or any text or logo. A photo moves only with a slow push-in (scale 1 to 1.04 over its scene).

Words (write them before you design anything)
- Write for a developer or product person scrolling past. Say what the article lets them do and why it matters, in its own concrete terms: its product names, numbers, steps and results. Cut any line that would fit a different article.
- The opening headline is the hook, never the article's title. Read in order, the headlines alone tell the story, each a short statement with a full stop ("One of the first to integrate it.", "Plain language in. A workflow out.").
- The voiceover sounds like an engineer explaining it to a friend: short sentences, "you", contractions, exactly one sentence per scene. Each scene's headline is the short form of its sentence: the same point and the same key words, so what's on screen and what's said tell one story.
- The sentences are one story in order: the hook said aloud (the first sentence opens with its words) and why it matters, how it works in two or three steps, the proof (the article's most striking fact), what you end up with, and last what the reader gets from the full article, specifically (never just "learn more"). Read the voiceover alone, aloud: it must make sense as one paragraph to someone who never saw the article, each sentence following from the one before. Rewrite any sentence that needs the screen to make sense, points back with "it" or "this" to something it never named, or reads like a fragment.
- Never write: unlock, seamless, effortless, revolutionize, game-changer, supercharge, elevate, leverage, harness, cutting-edge, next-level, "the power of", "in today's world", "dive in", or exclamation marks.
- Only state facts the article gives. Never make up numbers, quotes, names or claims.

Sound and captions
- Music: one audio clip generated with the elevenlabs-music model from the prompt "minimal ambient electronic, soft synth pads and light percussion, 120 BPM, instrumental only, no vocals". It starts at 0, lasts 32 seconds, at volume 0.18. Don't use a composition plan.
- Voiceover: generated with elevenlabs-multilingual-v2, 60 to 75 words in all, in the article's language, numbers spelled out, calm and warm: exactly one sentence per scene, in scene order, each ending with a full stop or a question mark and no other full stops (no abbreviations like "e.g.").
- Time everything to the voice, never to the clock, so each scene starts exactly when its sentence does: the voiceover is one audio clip per sentence, all on one track, with aliases "s1", "s2" and so on. The first starts at 0; every other one has start "auto", so it follows the one before. All have length "auto" and the same model and options, so it's one voice. Each scene's html5 clip, its photo and its captions take start "alias://s1" and length "alias://s1" (and so on) from their sentence; the last sentence plays over the end card. You only know each length at render: the voice speaks about 2.5 words a second, so finish each scene's build-up a second before (its sentence's words) / 2.5 seconds, then hold. Write those estimates into script.md next to each sentence before you build.
- Captions: one rich-caption clip per sentence, all on one track, each with src, start and length "alias://" of its sentence. One short phrase at a time, in sentence case: Inter at weight 600, 54 px, in the look's text colour, with the spoken word in the accent colour (white over the end card's accent background, where the accent would vanish). No background, no box or pill behind the words, and no outline. Each caption clip is 912 x 200 with fit "none", centred at y 1620 in the bottom zone, and nothing else enters the bottom zone. Where a full-bleed photo sits behind the band, the scene's html5 fades it into the look's background colour there with a soft gradient, so the words stay readable without a box.

Easy to miss
- tracks[0] is the top layer: the captions first, then the html5 scenes, then the photos, then the end card, then the audio.
- A rich-text animation style is "character" or "word", nothing else.

What’s in each short

  • A hook, not the post’s title. The video opens on a 3 to 6 word line that makes a scroller stop, and the voiceover says it. Under it is what the post is actually about, its product or the thing it makes, rebuilt on screen, not stock art.
  • One idea per scene, 7 or 8 scenes. Product-style screens, lists and one big number, rebuilt from the post’s real names and data.
  • Every scene starts when its sentence does. The voiceover is one AI voice clip per sentence, and each scene is pinned to its sentence, so what’s on screen always matches what’s said.
  • Captions transcribed from the voiceover, one phrase at a time, with the spoken word highlighted.
  • An end card that shows the post’s page, so viewers know where to read the rest.

No template: Claude designs every video

There’s no fixed layout. Claude follows the brief’s rules for the look (the type sizes, the zones of the frame, the colours, the words) and decides the scenes itself, from the post. [AFTER THE RUN: one or two sentences on what Claude made for the post in the video above.] Run the same post again and you get a different video.

How to set it up

It takes about 15 minutes. You need a paid Make plan, because a short takes about 30 minutes and Make’s Free plan stops a run at 5 (Core runs for up to 40).

  1. Get a Shotstack API key.
  2. Get an Anthropic API key with access to Managed Agents (beta).
  3. Make a folder in Google Drive for the shorts and copy its ID: the last part of the folder’s link.
  4. In Make, create a scenario and import the blueprint (⋯ → Import Blueprint).
  5. Add your connections: Shotstack on the two Shotstack steps, Anthropic Claude on every Claude step, and Google Drive on the two Drive steps. In both Drive steps, replace YOUR_FOLDER_ID with your folder’s ID.
  6. In the RSS step, replace Shotstack’s feed with yours, then click Run once.

What it costs and how long it takes

What Cost
Shotstack, per short About 6 credits: the music is 3.75 credits, each photo 0.75 (one or two), the voiceover about 0.05 a sentence and the 32-second render about 0.5
Claude Billed by Anthropic for the agent’s session. The scenario caps each session at $10, and you can change that in the step that starts the session
Make About 230 operations a short, so Core’s 10,000 a month covers about 40
Time About 30 minutes: Claude takes 18 to 25 to write and check the edit, and Shotstack 5 to 12 to render

Ideas

  1. Swap Run once for RSS → Watch RSS feed items and schedule it, so every new post on your blog becomes a short on its own.
  2. Point it at a client’s blog and hand them a Drive folder with a short for each post.
  3. Change the brief for other formats: a 60-second explainer, a square video for LinkedIn, or your own brand colours.

Good to know

  • The RSS step reads the newest post only. Keep it at 1: a short takes about 30 minutes and a Make run stops at 40, so a second post in the same run gets cut off.
  • The brief is plain English in the Transform to JSON step, right after the page is downloaded. Claude’s working instructions sit in the step that starts the session. Change the length, the look or the voice there.
  • If an AI asset fails to generate, Shotstack stops the render and you get the failed note. Run that post again; it usually works the second time.
  • I tested it with Claude Opus 5.5 at high effort.
  • Read the words before you post a video anywhere.

Links

If you try it, post your short in the thread, I’d love to see what it makes from your blog. And what should I build next?

Cool, for the rest, you can also run remotion free on your device and connect it to voice agents etc. to build videos. You can have make AI Agents do the scripting and brand research, etc.