You approve four times. We do everything else.
Most tools hand you a pipeline and call it power. This one runs the pipeline and asks you only where a human decision genuinely changes the outcome.
What you give us
One brief and a short checklist. That is the whole input.
One sentence
What the video is about. Not a script, not a shot list — a sentence. “A 40-minute Arabic documentary on the collapse of a famous bank.”Channels you like
Point at channels whose format works. Name one and the system adds four more by itself, because a formula measured from a single channel is that channel’s identity, not a format.Your identity
Name, language, voice, palette, the topics you will and will not touch. Given once, changed whenever you want, never overwritten by the system.
The machinery is not your problem
There is no template store, no style you have to pick from a gallery, no node graph. Topic selection, script, sourcing, shot planning, image generation, voice, assembly and quality control run without you.
You are shown what happened and why, at each of the four points where your answer actually changes something.
Where the pipeline stops and waits for you
These four are gates. Everything between them is ours.
Proof
Every factual claim becomes a row with a source, a tier and five integrity checks. Nothing downstream runs until zero flags are open.Plan & cost
The router assigns every shot to the cheapest method that can do it well, and the cost is shown before a single credit is spent.Packaging
Three title options and three thumbnail options, each with the reason it might work.Publish
The final human gate. Performance comes back and updates the Brain’s memory.
A job waiting on you says needs you, in those words, everywhere it appears. It never hides behind “pending”.
11 stages, four of them yours
Published in full because you should be able to see what you are buying.
Brief
One sentence from the customer plus the checklist answers.Topic
Candidates ranked by demand, competition gap, sourceability and channel fit. The winner is shown with its reasoning.Script
Written inside a locked contract and quota-checked against the Brain’s Formula. Never sees a blocked identity marker.Proof
Every factual claim becomes a row with a source, a tier and five integrity checks. Nothing downstream runs until zero flags are open.Plan & cost
The router assigns every shot to the cheapest method that can do it well, and the cost is shown before a single credit is spent.Visuals
Batched, with review gates. Failures are refunded automatically.
Voice
Chunked at beat boundaries, never mid-sentence, and regenerated when a chunk comes back short.Assembly
Edit, then the camera pass and the audio pass — both free, both what makes it look filmed.Quality check
17 automated checks. A blocker stops delivery.Packaging
Three title options and three thumbnail options, each with the reason it might work.Publish
The final human gate. Performance comes back and updates the Brain’s memory.
Every shot goes to the cheapest method that can do it well
Quality does not come from picking the most expensive model. It comes from not sending a bar chart to an image generator.
Accurate things are drawn in code
Text, numbers, charts, timelines, maps, quotations and citations are rendered as code, not generated. They are exact, identical every time, and they never misspell a word or invent a digit — which no image model can promise.
They also cost nothing to produce, which is why an episode can carry 167 visuals without the cost following the density.
Texture comes from an image model
People, places, atmosphere and grain come from an image model, asked for native 2K directly.
We never pay for an upscale. If you want 2K, ask the model for 2K — an upscaler is a second bill for something you could have had first time.
Most motion is a camera move
Roughly 90 to 95 percent of the motion in an episode is a slow push or drift across an approved still.
It costs nothing, and unlike a generated clip it never warps a face halfway through the shot.
Two to four beats get true motion
Real generated motion is reserved for the beats where movement changes how the moment lands: the hook, a reveal, a turn, the payoff. Never more than 4.
Each one starts from a still you already approved. A generated clip costs many times what a still costs and is rarely better, which is exactly why it is rationed rather than sprayed across the episode.
What makes it look filmed instead of rendered
Both passes are deterministic post-processing. Neither costs a model call.
The camera pass
- Film grain, matched to the scene rather than laid flat over everything.
- A little lens distortion, because a real lens has some.
- White balance that is slightly imperfect, the way a real camera is.
- A phone-like encode, so it sits naturally in a feed of phone footage.
The audio pass
- A room impulse response, so the voice sounds recorded in a place.
- Room tone underneath, because true digital silence is the tell.
- Loudness normalised to -16 LUFS, so your episode is not markedly quieter than everything around it.
17 automated checks before anything is delivered
A blocker stops delivery. A warning is shown and can be accepted. A check that did not run counts as a failure, not as a pass.
| Check | What it catches | Severity |
|---|---|---|
| Audio is actually there | the silent-track bug — a perfect-looking file with no sound | blocker |
| Loudness normalised | an episode that is markedly quieter or louder than everything else on the platform | blocker |
| Duration matches the script | a dropped beat, a truncated voice chunk, or a stuck last shot | blocker |
| No truncated voice chunk | a voice provider silently cutting a long chunk short mid-sentence | blocker |
| No black frames | a missing asset that renders as a hole in the episode | blocker |
| No frozen video | a Ken Burns move that failed to apply and left a static card | warning |
| Audio and video are the same length | drift that shows up as lip-sync failure at the end of a long episode | blocker |
| Resolution and aspect ratio | a vertical ad delivered horizontally | blocker |
| Constant frame rate | variable-frame-rate output that some platforms re-encode badly | warning |
| File size within the delivery limit | an upload that fails at the very last step | blocker |
| Captions inside the safe zone | text hidden behind the TikTok UI or the YouTube progress bar | blocker |
| Every claim shows its citation while it is spoken | a sourced script that ships as an unsourced video | blocker |
| Every claim has a source | a claim that slipped past the proof gate through a later edit | blocker |
| The rendered file says what the script says | the wrong voice take, a stale script version, or a mis-ordered chunk | blocker |
| Thumbnail readable at phone size | a thumbnail designed at full size that is mush in a feed | blocker |
| No blocked identity marker reproduced | the writer drifting onto another channel’s signature phrase | blocker |
| AI-content disclosure set | a policy strike on the customer’s own account | blocker |
The first check exists because of a specific failure: an audio track reading near −91 dB is not quiet, it is silent, and it ships as a perfect-looking file with no sound.
Honest ETAs, published
Render time is real work on real hardware. These are the numbers we plan against.
| Episode | Length | Typical render |
|---|---|---|
| Short | 5–8 min | 25 min |
| Standard | 10–14 min | 45 min |
| Long | 18–25 min | 1.3 hours |
| Documentary | 35–45 min | 2 hours |
- A 45-minute episode is about 2 hours of render. We would rather tell you that now than apologise later.
- Queue time is on top of render time and depends on how busy the queue is. It is shown live inside the product.
See the plan before you spend anything
Create a workspace, describe one video, and read the shot plan and its exact cost in credits. Approving it is a separate decision.