Five Of These Ten Prompts Name A Flicker, A Warp Or A Jump
EDIT VIDEOS WITH AI. NO TIMELINE NEEDED. Ten prompts, posted in the comments, for reframing, relighting, replacing backgrounds, removing people, adding holograms and turning a photograph into an advertisement.
That headline is the second time this week we have read it. Yesterday’s card made the same promise and kept it the opposite way – by writing the edit as code, so the render is a function of time and nothing is generated at all. This card keeps it by handing the footage to a model and asking it to repaint.
Same promise, opposite method. Which makes the two cards a useful pair, and makes one question the whole thing turns on: do your original pixels survive?
But start with what the author got right, because it is the most technically literate thing on the card. Read the defensive clauses:
- #6: “eliminate flickering, warping, duplicates, or visible artifacts“
- #8: “no jumps, freezes, flickers, or sudden changes“
- #4: “Match perspective, lighting, shadows, depth, color, and camera movement“
- #7: “matching perspective, lighting, reflections, depth, motion, and camera movement“
- #5: “no excessive smoothing, saturation, or HDR”
Five of the ten name a temporal fault by name. Flicker. Warp. Duplicate. Jump. Freeze. Not one of those words describes something wrong with a single frame – every one of them describes two frames disagreeing.
That is the whole difference between editing a photo and editing a video, and whoever wrote these knows it. A still image either looks right or it does not, and you can tell in a second. A ten-second clip at thirty frames a second has to look right and stay the same thing for three hundred consecutive frames. The subject’s jaw cannot drift. The new background cannot shimmer. The removed person cannot leave a ghost that pulses.
And you cannot see any of that in a screenshot.
The post this article is about
Read the original post: “Edit videos with AI. No timeline needed.”
All ten prompts are posted in the comments beneath. The tables below quote them as written, read on 12 October 2026.
Adjusting footage and inventing footage are different jobs
The ten prompts look alike – same structure, same preservation clauses, same bracketed blanks. They split cleanly on one question, and the split predicts almost everything else.
Diagram: adjust versus generate
| ADJUST your footage | GENERATE new footage | |
|---|---|---|
| Which prompts | 3 landscape → vertical · 5 fix bad lighting | 1, 2, 4, 6, 7, 8, 9, 10 — eight of the ten |
| What happens to your pixels | They are kept and altered | They are replaced by pixels that were never filmed |
| How likely it is to work | High. Reframing and grading are close to deterministic | Varies enormously. Hardest when the camera moves and the subject has fine edges |
| What fails | A crop you dislike, a grade that is too warm — visible immediately | Flicker, warping, a ghost, a shimmering edge — visible only in playback |
| Is the result still a record of what happened? | Yes. Same event, different framing and colour | No. You are now somewhere else, or something is missing, or something is there that was not |
| Disclosure on YouTube | Colour and lighting fixes are listed among the exemptions | Read the rules. “Alters footage of a real event or place” is one of the four triggers |
| How to check it | Look at it once | Scrub it frame by frame at the edit boundary |
Eight of the ten are in the right-hand column. That is not a criticism – generating is the whole point of the card, and the right-hand column contains everything exciting on it. It is a warning about where to spend your checking time, and about the fact that both columns are written in the same confident voice.
Row four is the one that costs people a published video. On the adjust side, a bad result is obvious the moment you look. On the generate side, the still frame you exported to check it can look perfect while the clip shimmers – and the still is exactly what most people check.
All ten, sorted by whether your footage survives
Column 2 is the card’s own wording. Column 3 is the sort that decides how much checking each one needs.
| Prompt | Quoted as written | Adjust or generate | Our note – read 12 October 2026 |
|---|---|---|---|
| 1. Professional Reel Upgrade | “Preserve the person, voice, expressions, movements, and dialogue. Replace the background with [BACKGROUND]… Keep the subject well-framed with clean space for captions.” | Generate (background) | “Clean space for captions” is the detail that gives away a working creator. Nobody thinks of that until they have had a caption land on someone’s chin. But note it both preserves the person and replaces where they are standing |
| 2. Cinematic Product Ad | “Preserve the product’s exact design, branding, colors, and details. Place it in [SETTING] with [LIGHTING]… finish with a powerful hero shot.” | Generate (setting) | The preservation clause is doing serious work here – an altered logo or a changed product colour is a commercial problem, not an aesthetic one. Check the branding frame by frame |
| 3. Landscape → Vertical | “Reframe intelligently around [MAIN SUBJECT] while preserving important people, products, text, and actions. Keep key elements within the safe area.“ | Adjust | The most reliable prompt on the card and the least glamorous. No pixels are invented – it is a moving crop. “Safe area” matters because the bottom of a vertical video is where the platform puts its own interface |
| 4. Background Replacement | “Keep the subject unchanged and naturally integrated. Match perspective, lighting, shadows, depth, color, and camera movement… Add subtle environmental motion such as [DETAILS].” | Generate | The hardest one on the card. Matting a moving subject with hair, glasses or fast motion is where this breaks. “Add subtle environmental motion” is clever – a perfectly still fake background is itself a tell |
| 5. Fix Bad Lighting | “Fix exposure, white balance, contrast, shadows, and highlights while preserving natural skin tones and facial detail… Keep the result realistic – no excessive smoothing, saturation, or HDR.” | Adjust | Genuinely safe and genuinely useful. This is colour grading, which is what every editor has always done. The “no excessive smoothing” clause protects the face from being quietly beautified |
| 6. Remove Unwanted Objects | “Remove [OBJECT/PERSON] and realistically reconstruct the background… eliminate flickering, warping, duplicates, or visible artifacts.” | Generate | Names four temporal faults in one clause – the most technically aware sentence on the card. Watch specifically for a faint ghost where the object was, pulsing as the camera moves |
| 7. Cinematic VFX | “Add a realistic [PORTAL/HOLOGRAM/ENERGY TRAIL/PARTICLES] at [MOMENT]. Integrate it by matching perspective, lighting, reflections, depth, motion, and camera movement.” | Generate | “Reflections” is the expert word. A convincing added object changes the light in the scene around it; an unconvincing one floats. Asking for reflections is asking for the hard half |
| 8. Seamless Loop | “Make the ending flow naturally into the opening… Hide the transition completely – no jumps, freezes, flickers, or sudden changes.” | Generate (the join) | A loop is judged at exactly one moment, which makes it the easiest thing on this card to check: watch the wrap point five times in a row |
| 9. Photo → Video Ad | “Turn this product image into a premium 9:16 cinematic ad… Preserve the exact product design, branding, colors, and details.” | Generate – everything but one frame | There is no footage here at all. One still becomes a moving ad, which means every frame except the source is invented. Highest reward, highest drift risk |
| 10. Idea → Social Video | “Create an engaging 9:16 social video about [TOPIC] for [AUDIENCE]… a strong curiosity-driven hook, three useful points… End with: ‘[CTA].’” | Not editing at all | This one does not belong on a card about editing. There is no input footage – it is generation from a topic. Good prompt, wrong card, and it is the one where the output is least yours |
Two ways to lose the timeline
Put this card next to yesterday’s and the contrast is clean, because they solve the same complaint with opposite philosophies.
The Claude card removes the timeline by making the edit deterministic: one HTML scene that draws any frame at time t, rendered frame by frame through a browser. Nothing is generated. Run it twice and you get the same file, byte for byte. Its selling line was “renders never drift” – and drift is impossible when every frame is computed from the clock rather than from the frame before it.
This card removes the timeline by making the edit generative: describe what you want and a model repaints the pixels. Run it twice and you get two different videos. Drift is not impossible here; it is the thing the five defensive clauses are fighting.
Neither is better. They are good at completely different jobs, and the honest summary is short:
Code wins when the output is structured – titles, data, templates, fifty variants of the same explainer. It cannot invent a forest behind you.
Prompts win when the output is photographic – a new location, a removed stranger, a product in a studio you do not have. No amount of HTML will produce that.
And the costs mirror each other exactly. The code route costs you a toolchain – Remotion, Playwright, ffmpeg, five dependencies – and gives you reproducibility for free. The prompt route costs you nothing to set up and charges you on every single clip, in checking time.
Checking video is not like checking an image
Here is the practical half, and it is the thing the card leaves out entirely: there is no verification step anywhere in these ten prompts. The checks below take two or three minutes and they are deliberately different from the ones you would run on a photograph.
| Check | How | What you are looking for | Which prompts need it most |
|---|---|---|---|
| 1. Watch it at 2× | Play the whole clip at double speed, once | Flicker becomes obvious when sped up and nearly invisible at normal speed. Thirty seconds and it catches most of it | All eight generate prompts |
| 2. Scrub the edges | Step frame by frame where the subject meets the new background. Hair, glasses, fingers | Shimmering outlines, a halo, strands that appear and vanish | 1, 4 background replacement |
| 3. Hunt the ghost | Go to where the removed object was and step through slowly | A faint patch that breathes as the camera moves. The reconstruction is per-frame guesswork and the guesses disagree | 6 remove objects |
| 4. Read the text and logos | Pause on any frame containing words or branding, then the next, then the next | Letters that change shape between frames. Text is where warping shows first | 2, 9 product ads |
| 5. Watch the wrap point | Loop it five times and stare at the join | A jump, a freeze, a lighting shift at the seam | 8 seamless loop |
| 6. Check the face against the original | Same frame, before and after, side by side | A jaw that narrowed, skin that lost texture. Relighting is where quiet beautifying hides | 1, 5 |
| 7. Ask what it would look like if it were wrong | Before you export: name the one thing most likely to have broken, then go and look at that | Checking without a hypothesis is just watching your own video and liking it | Everything |
Check one is the cheapest thing on this page. Playing a clip at double speed makes temporal inconsistency jump out, because your eye stops tracking the content and starts noticing the wobble. It costs thirty seconds and catches more than any amount of careful frame-staring.
Check seven is the habit. It is the same move the Claude card built into its own pipeline as step 6 – review stills from your own render and fix what is wrong – except here nothing does it for you, so you have to.
Four honest caveats
These are well-written prompts and the temporal clauses prove it. Naming flicker, warping, duplicates, jumps and freezes is not something you do unless you have watched an AI edit fall apart in playback. Prompt 7 asking for reflections and prompt 4 asking for subtle environmental motion are both expert touches – they are the details that separate a composite that holds from one that floats.
The disclosure question is sharper here than on most cards. We have read YouTube’s page on disclosing GenAI content twice while writing this series. Two of its four triggers are “alters footage of a real event or place” and “generates a realistic scene that didn’t actually occur” – which is a fair description of prompts 1, 4, 6 and 9. The page also lists colour and lighting fixes among the exemptions, which covers prompts 3 and 5. And it says in writing that disclosing “won’t limit a video’s audience or impact its eligibility to earn money”, so the box costs nothing to tick.
Prompt 6 deserves a separate thought. Removing a person from footage is the one entry here with an ethical edge as well as a technical one. Removing a stranger who wandered through your shot is housekeeping. Removing someone who was part of the event, from footage you present as a record of that event, is a different act – and no prompt clause covers the difference.
And we have not run them. Everything above is read off the prompts as written, plus what is generally true of video generation. We have not measured how any particular model handles any particular clip, and results will differ wildly by footage – a locked-off shot and a handheld walk-and-talk are not the same problem.
The one-line version
Five of these ten prompts name a flicker, a warp, a jump, a freeze or a duplicate, and that tells you exactly where AI video editing fails: not inside one frame, but between two. A photograph either looks right or it does not. A clip has to look right and then stay that way for hundreds of consecutive frames, and nothing you can export as a still will tell you whether it did.
So sort the ten by one question – do my pixels survive? Prompts 3 and 5 adjust your footage and are nearly safe. The other eight invent pixels that were never filmed, and they carry all the risk, all the excitement and the whole of the disclosure question.
Then check by playing, not by looking. Watch it at double speed once – thirty seconds, and flicker that is invisible at normal speed becomes impossible to miss.
And if you take one prompt: take number 3. Landscape to vertical is the least glamorous entry on the card, invents nothing, and is the one most creators need every single week.
Learning a lot does not make anyone good. Doing a lot does. Ten prompts is one evening of experiments; one clip you actually checked frame by frame before publishing is the habit that keeps the channel.
If you want to be held to that, we run a free challenge at BANI Academy. There is no barrier to entry – you commit a small fee that is refunded when you finish, and the only requirement is that you actually do the work for the length of the challenge.
Run one of these, then watch the result at 2× and tell us what you caught. Comments are under the original post – the artifact you found, not the prompt you used.
More on building a channel for a global audience at mmoyoutube.com.
