Most AI content tools have a dirty secret: nobody, including the founders, can say exactly what the AI will produce. The model writes the copy, picks the layout, times the cuts — and when a video comes out weird, there's no bug to fix, just a prompt to nudge.
We're building Man Friday the opposite way, and last week we ran the test that proves the architecture: produce three real, postable videos with zero AI calls in the render path. Slots filled by hand, everything downstream plain code.
The rule: models write words. Code makes videos.
Our pipeline confines the LLM to exactly two jobs: reading your product's website into a brand brief, and filling small text slots — a hook, five slide lines, a caption — inside a template that a human curated from a video that already went viral.
Everything else is boring, deterministic code. The template declares the structure: slide count, timing, layout positions, style tokens, where your product screenshot goes, which backgrounds are allowed. A render worker executes it with an image library and ffmpeg. Same inputs, same video, every time. A weird video is a bug you can find.
Why hand-produce three posts first?
Because a spec you haven't used is fiction. We wrote the slot text ourselves — no model — for one slideshow, one faceless hook video, and one avatar script, then rendered all three. The exercise caught three real spec bugs:
- Our spec failed its own example. The lint rule said every declared slot must appear in the render plan — but the caption slot is consumed at publish time, not render time. Every valid template failed. The spec now distinguishes render slots from publish slots.
- Video timing was undefined. Voiceover length varies, so who decides how long each shot holds? The fix: one "flex" shot per template absorbs the voiceover's real duration. Templates never author a timeline, and sync falls out for free — the script's product mention lands exactly as the CTA card cuts in.
- Nothing said what happens when an asset is missing. Stock footage now falls back to a brand gradient instead of failing the render.
None of these would have surfaced in a slideware demo. All three would have surfaced in production, at 2am, as a customer's broken video.
The part users actually feel
Determinism isn't the pitch — it's the cost structure that pays for the pitch. Because a concept re-renders forever without touching a model, we can pre-generate your feed overnight in a cheap batch, keep previews instant, and spend real money only on the videos you approve. The experience gets the budget; the pipeline runs on arithmetic.
The template spec, the lint, the renderer, and the three test posts are all in our repo's validation folder. More of these notes as we build — the pipeline goes headless-first over the next three weeks, and we'll publish what the first ten generated concepts look like, good or bad.
Man Friday posts to TikTok and YouTube Shorts for solo builders — you approve with a swipe.