Generative video · TPS - The Production Playbook for Youtube
Your AI footage looks fake. Here is exactly why.
Eleven chapters of failures that were paid for with real generations. The three causes of the plastic look. Why rewording a rejected prompt never works. The four physics failures you can predict before you spend anything.
- 11 Chapters
- 36 Prompts
- 5 Scripts
- 4 Templates
Instant download · PDF, prompts and working scripts
Look inside
Eleven chapters
- 01The format is the product
- 02Why it looks fake
- 03The twenty second loop
- 04The moderation wall
- 05Physics the model gets wrong
- 06Animals
- 07Assembly
- 08Cost discipline
- 09Long form
- 10The prompt library
- 11Failure catalogue
The complaint is always the same, and always vague
It looks AI. Broken down, it is almost never the subject that gives it away. It is the camera, or rather the absence of one. Real footage is a record of a physical device failing to be perfect. Generated footage, left alone, renders an idealised view from nowhere.
Nothing described a camera
No camera in the prompt means a render, not a photograph. Floating viewpoint, perfect exposure, no noise. Everything correct and nothing believable.
Flattering light
Warm, soft, low-angle light is the beauty-retouch look. Ask for harsh sun or flat overcast and a great deal of the artificiality disappears.
One specific word
There is a single word almost everyone puts in their prompt that actively pushes the output toward looking synthetic. Removing it beat every other change tested.
What you actually get
Findings, not tips
Every one of these came out of a generation that failed, was diagnosed, and was fixed. Usually across several attempts, always at cost.
Rope follows the camera, not the load
A prompt stating four separate times that the rope must run upward still produced one running downhill into mid-air. Prompt language cannot fix it. Camera direction can.
Large water renders as dust
Five separate causes, and you have to correct all of them at once. One word does most of the work, and it has no dust equivalent.
Rewording a rejection never works
The classifier reads the situation, not your language. Once you accept that, you stop burning generations on rewrites and start restructuring the shot.
Animals are the strongest tell
Gait and weight expose the model instantly. The fix is counter-intuitive: show less animal, less clearly, for less time, filmed worse.
Structures collapse straight down
Any directional failure has to be specified positively and negated by name, or you get a pancake collapse every time.
The API trap with no undo
Unknown parameters are silently accepted and charged rather than rejected, and there is often no cancel endpoint. Probe with invalid values, never unknown keys.
In the download
A system, not an ebook
11 chapters, PDF. Format, realism, loops, moderation, physics, animals, assembly, cost, long form.
Six subject areas, ordered by hook strength, each carrying the specific failure that subject produces.
Loop, multi-clip sequence, danger split, plus the realism and audio blocks you paste verbatim.
Colour matching, loop closing, titling without drawtext, assembly, upscaling. Tested end to end.
Proof
The system, running in public
Every rule in this playbook came out of a channel that ships. These are unedited platform analytics from that channel, generated with the prompts and assembled with the scripts included here. Revenue figures are the only thing blanked out.
Built for you if
- You already generate clips and they keep coming back subtly wrong
- You are building a channel and want a format rather than one-off experiments
- Your prompts get rejected and rewording is not getting you through
- Your clips look fine alone and wrong when cut together
Not for you if
- You have never generated a clip. Start there first, then come back
- You want prompts to paste without understanding why they are shaped that way
- You are looking for a tool tutorial. This is about failures of the model class, not one app
Questions
Which tool is this for?
None specifically, and that is deliberate. The failures described are failures of the class of model rather than of one version, which is why the fixes have kept working across three generations of tooling. The prompts paste into any current text-to-video model.
Do I need to code to use the scripts?
No. They are command-line scripts you run with one line each, and every one prints its own usage. You need ffmpeg and Python installed. The README walks through a first run.
Is this just a list of prompts?
The 36 prompts are included, but they are the smaller half. The playbook is about why a prompt is shaped the way it is, so you can write your own for subjects nobody has covered.
What format is it?
A PDF plus plain text prompt files and Python and shell scripts, delivered as an instant download. Nothing to install, no account, no subscription.
Will this stay current?
The findings are about how these models fail structurally rather than about a version number. Rope direction, moderation behaviour and collapse direction have behaved the same way across every model tested.
Instant download
Stop paying for the same mistakes twice
$37
One payment. PDF, 36 prompts, 4 templates and 5 scripts.