I don't make videos. My AI pipeline does.

You build things. It came out great. Now what? Getting attention for what you built is a separate job. Different skills. Different time. Different energy. Even if you have all three — skills, time, energy — the harder problem is: most of what gets built doesn't reach anyone. The obvious fix is to have an AI make the video for you. That produces slop. Style drifts from scene to scene. Pacing is random. You can't iterate without the whole thing regenerating into a different video. This is spoolcast. It's an architecture for making real short-form illustrated video. Built for people who build things and don't want marketing to be a second job. In the next seven minutes, we'll walk through the pipeline that made that clip for under three dollars. And the agent layer that runs it in the background while you keep building. The fix isn't a smarter AI. It's a division of labor. The generic approach asks one model to do everything. Story. Script. Images. Animation. Camera. Timing. Rendering. When anything's wrong, there's no atomic unit you can change and re-run. That's why the output feels slopped together. Structurally, it was. In spoolcast, there's exactly one atomic driver. The script. Specifically, a shot list — a spreadsheet with one row per sentence the narrator will say, grouped into scenes. Change one row. Only the artifacts downstream of that row get rebuilt. Everything else stays untouched. Every other layer has exactly one job. And one rule. It cannot improvise. Here's what specifically keeps the script from being slop. One. Images are grouped by chunk, not by beat. A chunk is one to six sentences that share a visual moment. One illustration covers the whole chunk. Saves the budget. Keeps the eye from whipping around. Two. The image doesn't illustrate a sentence. It shows the chunk's narrative throughline. A chunk about ad saturation gets one image of an ad-wall. Not a picture of every individual word. Three. Chunks have a hard cap around fifteen seconds. Past that, the static image goes dead. Split it. New image. Four. Every sentence has to survive on its own. If a line only makes sense when you read the one before it, that's an essay. Rewrite until each beat can be lifted out and still land. Five. Text-to-speech reads exactly what you type. So you write it the way you want it said. Write 'roe-ass.' Not 'ROAS.' Otherwise it says each letter out loud. Like a roll call. Six. One stage in the pipeline is deliberately not automatable. Turning messy raw material into a script needs judgment. What's the practical question. What's the turning point. What's the one thing the viewer should come away with. Mechanize that and you get generic. Leave it human. The rest of the pipeline takes over from there. Before the layers. The economics. Per five-minute video. Images — about fifty illustrations on kie.ai — roughly one to three dollars. Voice — Google Cloud's Chirp3-HD text-to-speech. Free within the monthly tier. Animation and render — both run locally. Zero cloud cost. Total per video: roughly the price of a coffee. Hiring a human animator for a five-minute explainer runs five hundred to two thousand dollars. Adobe plus Descript plus stock footage — fifty a month in subscriptions and a dozen hours of manual work. Spoolcast is an order of magnitude cheaper than either path. Four layers. Image makes the pictures. Animation gives them motion. Voice narrates. Render stitches it all into an mp4. Each does one thing. Together they make the video. Image layer. One AI illustration per chunk. The service is kie.ai. Their nano-banana-2 model. Each image is one HTTP request — a JSON body with a prompt and an image_input array. The problem this solves. If you ask a model for fifty illustrations, you get fifty art styles. Scene one is a cartoon. Scene two is a watercolor. Scene three is photorealistic. The fix is called image-ref chaining. First generation establishes a style anchor. Every subsequent request passes that anchor back in. The model matches the existing style instead of inventing a new one. Fifty scenes. One artist. Animation layer. A still image held on screen for eleven seconds while someone talks is dead air. The image needs to appear over time. The naive fix is a CSS wipe. Robotic. Up-down or left-right. Obvious every time. The real fix is a Python script that computes a per-pixel reveal-time map. For every pixel, pick a time between zero and the chunk's duration when it should become visible. Connected shapes reveal in parallel. Big shapes start early. Small details fill in later. Everything finishes together. Looks like an artist speed-painting. Not a computer wiping. The script uses OpenCV, a computer vision library. Runs locally. No AI tokens. No network calls. No randomness. Same image in, same frames out. Always. Voice layer. Google Cloud Chirp3-HD text-to-speech. The voice called Puck. Played at 1.2x speed. Default speed sounds like a robot. 1.2x sounds like a real podcast host. That's the whole trick. The earlier pick was ElevenLabs. Marginally better. Hundreds of dollars a month at daily-video pace. Chirp3-HD is free at this volume. Close enough nobody notices. Render layer. Remotion. A React-based video framework. Each scene is a React component that renders a frame. Running one command — npx remotion render — produces an mp4. Headless. No browser opens. It just runs. Deterministic. Same inputs, same mp4. Always. Four layers. Each doing one narrow thing. Nothing improvises. Here's what the whole pipeline produced. Five minutes. Forty-four chunks. AI voice. AI images. AI-written title. AI-generated thumbnail. Zero overlay improvisation. Rendered headless. One command. One thing left — and this piece is still being built as of this video. Someone has to decide there's a video worth making and write the brief. The plan is to delegate that too. The design is an agent that lives in your chat room. In this project — Zara. An Animabot running on a Matrix server. Built and talking in chat already. Not yet wired up to spoolcast. Once it's wired up, Zara will watch every signal the builder generates while working. Commits pushed to GitHub. New tracker entries. Chat messages about a milestone. A thread where the builder explained what they just figured out. When the signals line up, Zara writes the brief. A few paragraphs. Project name. Practical question. Turning point. Core message. The brief goes to the editorial agent. The editorial agent returns a shot list. Spoolcast renders. Zara drops the finished mp4 back in the chat with a short note. From the builder's side, once this is live, the whole thing will feel like this. You ship a thing. You go back to building. A few hours later, a video about what you just built appears in your chat. You won't write the script. You won't generate images. You won't edit anything. You did the building. The content got made. The 'mostly' in 'mostly automated' is the editorial agent that still needs real context. Because the alternative is a generic summary nobody watches. Mostly automated. With one intentional judgment layer that keeps the output from being slop. Using AI video to get attention used to mean one of two things. Producing slop. Or spending a week on every video. It doesn't anymore. Scope the AI narrowly. Let deterministic code handle the rest. Put an agent in the chat room to pull the trigger. The builder keeps building. The content gets made. Passively. And mostly automated. Spoolcast. The rendering engine for passive content.