Carousel Engine
Get the repo
An automated carousel system for AIVM
Write the copy, pick a layout, and finished carousel slides come out the other end, on brand every time.
Why this exists
Look at the public accounts and the pattern is hard to miss. The crypto side posts often and gets read. The enterprise side posts rarely and gets ignored, and the difference is not the quality of the thinking. It is that nobody has time to make the assets.
That is a throughput problem, not a design problem, which is why the answer here is a system rather than a set of posts.
The work
Try it
Pick any slide above and it opens here. The editor runs the same code as the command line, across all three engines, so a slide from the earliest system edits exactly like one from the newest. Anyone on the marketing team can drive it, and nothing they touch can take a slide off brand.
Live. Every edit changes the same JSON the command line reads.
How it works
A post is a JSON object: copy, layout, ground, object placement. Versioned, diffable, and reviewable before anything renders.
{ "kind": "poster", "bg": "tile",
"head": "How agent actions<br>flow through <b>AIVM</b>",
"obj": [{ "of": "flow", "opacity": 0.92 }] }
Every colour and size lives in one brand.json, authored against a 1080px frame and scaled
per format. A story crop is not a second design.
Codex generates a 3D render on black. The converter turns it into a field of cells by tracing its edges, which is what makes it read as line art rather than a silhouette.
Headless, at 2x, straight to PNG. Same output every run, so a post can be regenerated months later and come back identical.
node src/render.js --only aivm-matrix → output/aivm-matrix/portrait/*.png
What it produces
Portrait 1080×1350 is tuned and shipping. The renderer also emits square and story from the same spec, but object placement is authored against the 4:5 frame, so those crops still need per-format tuning before they are worth posting. One flag, then a pass of real work.
--formats portrait,square,story
Distinct page architectures, not variants of one block.
If a slide needs a form the library does not have, one command generates it, converts it, and prints the line to paste. It uses whichever generator you have: an OpenAI key, the Codex CLI, or any render you hand it as a file.
node src/gen-object.js --name shield \ --brief "two curved overlapping plates"
Every factual claim traces to aivm.io or brain.aivm.io and carries the date it was checked. A claim that is not in the registry, or a phrase on the blocked list, stops the render before a file is written. Nothing reaches a slide that has not already been verified against the source.
Install
Two pastes. The first sets it up, once. The second is what anyone on the team types every time they want a carousel.
Clone https://github.com/Arnav23apr/carousel-engine into a folder and run `npm run setup` inside it.
Paste into Claude Code. It clones, installs dependencies, fetches the browser the renderer needs, and drops the skill file where Claude looks for it. Once only.
Make me a five slide AIVM carousel about runtime policy enforcement. Use the ledger layout for the proof beat, and show me the PNGs when you are done.
Every time after that, this is the whole interaction. Claude reads the skill file, writes the spec, renders it, and hands back finished slides. Swap the topic for anything.
One command generates it, converts it to a cell field, and prints the line to paste into the spec. It uses whichever image generator the team already has, picked automatically:
OPENAI_API_KEY. Nothing to install.
--from render.png.
node src/gen-object.js --name shield --brief "two thick curved plates overlapping to form a guard"
Whichever route runs, the same style contract goes in and the same converter runs after, so an object made three different ways still reads as one library.
git clone https://github.com/Arnav23apr/carousel-engine cd carousel-engine && npm run setup
Node 20 or newer. Generating new objects also wants Python with Pillow and numpy; setup tells you if it is missing.
cd core node src/render.js --only aivm-matrix
Output lands in output/<id>/portrait/. Every sample in the picker above renders the
same way.
This is a prototype, not a finished system, and nothing here has been approved by anyone at AIVM. The display face is Outfit standing in for licensed Articulat. Several rows in the access log and two card headings are adapted from aivm.io, and the rest of the copy is mine and unreviewed. The browser export rasterises client-side for convenience; the command line produces the shipping files. What it is meant to show is the shape of the thing and how much of the carousel pipeline can be automated.