Create wants a complete scene. Edit wants a description of the change, not the scene. Motion wants environment only — motion verbs there cause jitter, because the movement already came from the driving video. Make a Custom Video wants a scene prompt plus five layers. Talking Actors wants a script, not a prompt. URL → Ad has no prompt field at all.
Almost every prompting guide assumes there's one box and one right way to fill it. There are five boxes here, and a sentence that's correct in one is actively wrong in another.
What does each field expect?
Five fields, and the differences aren't stylistic.
| Field | What it wants | What breaks it |
|---|---|---|
| Creative Studio, Create | A complete scene — subject, action, camera, light, tone | Chaining four actions into one prompt. One action per generation |
| Creative Studio, Edit | A description of the change only | Describing the whole scene again. The source already supplies it |
| Creative Studio, Motion | Environment, background, lighting, tone | **Any motion verb.** The driving video supplies all movement, and conflicting verbs cause jitter |
| Make a Custom Video | A scene prompt, plus five layers set separately | Leaving layers unset on a series. They get filled rather than skipped |
| Talking Actors | A script — the actual words spoken | Writing direction instead of dialogue. Duration comes from the word count |
| URL → Ad | Nothing. There is no prompt field | Expecting one. The link is the sole content input |
Source: Hexcoded Creative Studio, Marketing Studio, Talking Actors and URL → Ad documentation, checked September 2026.
A prompt isn't one thing. It's whatever the field is conditioned on, and that changes per field.
Create — the one every guide is about
This is the standard case, and the standard advice applies: be specific, one action per generation, name the camera move rather than calling it cinematic.
Two things worth adding that the general guides don't.
Prompt length is free. Only clip length, resolution and actor selection drive cost. There's no economic reason to compress a brief, and compressing one costs precision.
Character limits differ per model. Wan 2.7 and Vidu Q3-Turbo both show a 5,000-character counter; Hailuo 2.3 shows 2,000. Not every model displays one, so if you write long, check before you paste.
Edit — describe the change, not the scene
This is the field people get wrong most often, and the mistake is understandable.
Edit takes a source clip of three to ten seconds plus a description. The clip supplies the scene — framing, light, identity, everything already on screen. Your description supplies only what should be different.
So "a woman in a green jacket at a kitchen window, late afternoon light, she turns toward camera" is a Create prompt. In Edit it's noise, because the model already has all of that from the source and you've just given it competing instructions.
The Edit version is "change the jacket to navy." That's it.
Where this falls short. Three things can't be changed this way at all — identity, lighting and performance. Each is a property of the whole take rather than an element in it, so a described change alters more than you asked for. Those go back to regeneration.
Motion — the field that punishes normal prompting
Motion transfers movement from a driving video onto a character image. The movement is already decided by the time you reach the prompt.
Which means the prompt field is for environment, background, lighting and tone only. Hexcoded's documentation is direct about the consequence: conflicting motion verbs in the prompt can cause visual jitter.
So the instinct every other guide trains — describe the action precisely, add camera movement, be specific about what happens — produces a worse result here. You're giving the model a second opinion it didn't ask for.
The test is simple. Read each sentence and ask whether it describes the space or the person. If it describes the person, cut it.
A dim apartment kitchen at night, single overhead bulb, cool blue light from a window behind. Worn wooden surfaces, cluttered counter. Muted colour, slight grain.
Environment only. Nothing in this describes what the character does — the driving video carries all of it.
Make a Custom Video — prompt plus layers
Here the prompt is one input among six.
There's a scene prompt, and then five layers set separately: product, actor, format, hook and setting. Anything you leave unset resolves through Director's Choice, where the engine fills it against your prompt rather than omitting it.
That behaviour is useful on a one-off and wrong on a series. An unset layer is exploration when you're exploring and drift when you're on episode forty — because it resolves against this prompt, not against your previous episodes.
The same free-specificity rule applies, and more strongly. Prompt length, layer phrases, format, hook and setting selections and inline tagging all cost nothing.
Talking Actors — a script, not a prompt
The field takes the words your actor says. Not a description of them saying it.
Duration is derived from the script rather than set, capped at sixty seconds, with a live word count and estimate updating as you type. Roughly six words reads as three seconds.
Which means length is a writing decision rather than a setting, and there's no way to compress a long script into a short clip.
One control has no equivalent in any prompt field: inline emotion triggers. Manual emotion control takes a forward slash at any point in the script editor, placing a beat at an exact word rather than describing intent. For a clip written to a tight length that's how you buy back seconds — place the pause rather than adding words to signal it.
URL → Ad — no prompt at all
Three steps: paste a link, pick a format, choose an actor or faceless.
There's no prompt field and no manual script entry. The link is the sole content input, so everything the ad says about the product comes from what's on that page.
That makes page copy the prompt, in effect. A page with detailed specifications produces a specific ad; a thin page produces a generic one, and there's nowhere to compensate.
Which is worth knowing before you use it: the writing that determines the output happened on your product page, possibly months ago, possibly by someone else.
On URL → Ad, your product page is the prompt. Whoever wrote it wrote your ad.
How to keep them straight
Four questions, asked in order, before you write anything.
Does the model already have the scene?
In Edit and Motion, yes — from the source clip or the driving video. Describing it again competes with what's already there. Only describe what you're adding or changing.
Where is the movement coming from?
In Create you supply it in words. In Motion it comes from the driving video, so words describing it conflict. That single difference is the most common Motion failure.
Is this a prompt or a script?
Talking Actors takes the words spoken. Everything else takes a description. Writing direction into a script field just gets it read aloud.
Is there a prompt field at all?
URL → Ad doesn't have one. If the output is wrong, the fix is on the product page rather than in the tool.
- Create wants a complete scene. One action per generation, and name the camera move rather than calling it cinematic
- Edit wants the change only. The source clip already supplies the scene — describing it again competes with it
- Motion wants environment only. Motion verbs conflict with the driving video and cause jitter
- Make a Custom Video wants a scene prompt plus five layers. Unset layers get filled, not skipped
- Talking Actors wants a script. Duration comes from the word count, and inline triggers place a beat at an exact word
- URL → Ad has no prompt field. Your product page is the prompt
- Prompt length is free everywhere. Only clip length, resolution and actor selection cost credits
- Character limits differ per model — 5,000 on Wan 2.7 and Vidu Q3-Turbo, 2,000 on Hailuo 2.3
Because each field is conditioned on something different. Create has only your words. Edit has a source clip that already supplies the scene. Motion has a driving video that already supplies the movement. A sentence that adds information in one is competing with existing information in another.
Only the change. The source clip supplies framing, light and identity, so re-describing the scene gives the model competing instructions. "Change the jacket to navy" is an Edit prompt; a full scene description isn't.
Probably motion verbs in the prompt. All action dynamics come from the driving video, and Hexcoded's documentation states that conflicting motion verbs can cause visual jitter. Read each sentence and cut anything describing what the person does rather than what the space looks like.
No. Prompt length is free in every tool. Only clip length, resolution and whether you select a real human actor drive cost — so compressing a brief saves nothing and costs you precision.
It varies by model and not all of them display a counter. Wan 2.7 and Vidu Q3-Turbo both show 5,000 characters; Hailuo 2.3 shows 2,000. If you write long, check the counter before pasting rather than after.
Through the product page, since the link is the sole content input. A page with detailed, specific copy produces a specific ad. If the output is generic, the fix is on the page rather than in the tool — and Plan First will show you what it read before you spend anything.
Six fields, one library
Every tool draws on the same references and the same credit balance, so a character you saved in Creative Studio is castable in Talking Actors. What changes is what each field expects you to write.
Open Creative StudioMore on prompt structure, per-tool syntax and common mistakes in Prompts.