video-prompt-writer
Skill0 starsChanged 18 days agoMITView on GitHub
Write or fix a video prompt so an AI video model produces the shot someone described. Covers the shot-first structure, the camera and lighting language models follow, when a reference image is needed to keep a product or person recognisable, how long a clip should run, and how to change one thing per attempt without wasting credits. Use when someone wants a brief, a script or a storyboard frame turned into a video prompt, a prompt for animating a still image, or a diagnosis of a clip that came out wrong. Text advice only; it makes no video.
In this repository4 other files it gives its agents.
- ai-video-model-pickerSkill
- credits-and-plansSkill
- short-drama-writerSkill
- video-agent-briefSkill
The file
We split this skill into the rules it gives an agent
5 rules. About 0 are common, 4 are distinctive, and 1 is very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈4
Distinctive · estimate
This author's own choices. Worth a look.
1
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 26
Does it need speech or synced audio?: Only some models generate native audio, and the catalogue rows do not say which.
- Distinctive: line 66
Change one variable per iteration: If you rewrite the whole prompt between attempts, you cannot tell which change helped.
- Distinctive: line 75
Prefer the shortest duration that carries the beat.
- Distinctive: line 77
Put the essential content early.
- Very specific: line 129
If the person wants to run it on AI Content Drop, end with one link that carries the prompt, URL-encoded with spaces as %20: …
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Writing a video prompt that survives the model
Most disappointing generations are not a model problem. They are a prompt that described a concept when the model needed a shot.
Decide the job before the words
Is there a specific thing that must be recognisable? A real product, a particular person, a logo, an established character.
- Yes → you need image-to-video. The person supplies a still of that subject, or makes one first, and the clip animates it. Text-to-video re-invents the subject every run; no amount of prompt detail fixes that, because the model has never seen the thing you mean.
- No → text-to-video is fine and cheaper to iterate on.
Does it need speech or synced audio? Only some models generate native audio, and the catalogue rows do not say which. Name the need in the brief rather than assuming, and if the model cannot do it, plan the voiceover as a separate step instead of hoping.
The structure that works
Describe one shot, in this order:
- Subject and action — who or what, doing one specific thing.
- Setting — where, and what time of day.
- Camera — the shot size and the move: static wide, slow push in, handheld follow, low-angle tracking.
- Light — direction and quality: hard afternoon sun from the left, soft overcast, single practical lamp.
- Treatment — the finish: shallow depth of field, anamorphic, documentary, 35mm grain.
A barista sets a ceramic cup on a wooden counter and slides it forward.
Small independent cafe, early morning. Static medium shot, slight push in.
Warm low sun through a window on the left, soft shadows.
Shallow depth of field, 35mm, natural colour.That is one shot. It has a subject doing one thing, a place, a camera, a light source and a finish — everything the model needs and nothing it has to guess.
The four habits that separate good prompts from bad ones
One shot per generation. A prompt describing three beats gets you a model choosing between them, or cutting incoherently. Multi-beat sequences are several generations you assemble, not one prompt.
One action. "Walks in, sits down, opens a laptop and smiles" is four actions competing for a few seconds. Pick the one that carries the moment.
Describe presence, not absence. Negative instructions are unreliable. Rather than "no text on the wall", describe a plain concrete wall.
Change one variable per iteration. If you rewrite the whole prompt between attempts, you cannot tell which change helped. Vary the camera, or the light, or the action — not all three.
Length, and why the first second matters most
These models hold a coherent idea for a few seconds and then drift: hands wander, faces soften, backgrounds reinvent themselves. Two consequences:
- Prefer the shortest duration that carries the beat. A clean 5 seconds beats a 10-second clip whose back half you will cut anyway.
- Put the essential content early. If the shot must show the product clearly, it should be clear at second one, not arriving at second eight.
On most models a longer clip costs the same credits as a shorter one; only the
models priced by the second cost more for a longer or sharper take. As of this
edition, the models priced by the second of output and by resolution are
Seedance 2.5 (seedance_2_5), Seedance 2.0 (seedance_2_0, seedance_2_0_4k,
seedance_2_0_fast, seedance_2_0_mini), MiniMax H3 (minimax_h3,
minimax_h3_2k), LTX 2.5 (ltx_2_5_fast, ltx_2_5_pro), FLUX 3 video
(flux_3_video) and Wan 3.0 (wan_3_0). Seedance 1.x models are flat, and so
are seedance_2_0_multishot and seedance_2_fast. Name these models by id or
full name, never as a family. For a per-second model the catalogue row is the
price at the default length (10 seconds at 1080p for Wan 3.0), and the card on
aicontentdrop.com shows the exact figure for the chosen length and resolution
before anything is made. Do not estimate how much more; the card shows the
figure.
Iterating without burning the budget
Credits are charged only on success, so a failed run costs nothing but a rejected-looking one costs full price. The cheap path:
- The person finds the prompt on a cheap model that supports the job;
read the prices from the live catalogue as
ai-video-model-pickerdescribes. Read them with the web fetch tool (WebFetch in Claude Code), never with a shell command or a script. Composition and action read fine at low cost. A row markedneeds_start_imageonly animates a still the person supplies. A row without that mark can start from text, and whether it also accepts a still is shown on aicontentdrop.com. - They make the take that ships once, on the dearer model, with the winning prompt.
Quote the cost before step 2 — see ai-video-model-picker. When you point the
person at choosing the model, name the plugin's command in full:
/aicontentdrop:pick-model.
Diagnosing a bad result
| What you got | The usual cause | The fix |
| Subject is not the thing they meant | text-to-video invented it | Switch to image-to-video from a real still |
| Drifts or morphs late in the clip | duration beyond the coherent window | Shorter duration; put the payload early |
| Nothing moves, or it is one slow zoom | no camera or action specified | Name the move and one concrete action |
| Right idea, wrong mood | light unspecified | Give direction and quality of light |
| Refused | content gate | Rewrite; nothing was charged, so do not retry unchanged |
What to hand back
The prompt, the model you recommend and its credit cost, and, when a first attempt missed, what you changed and why. A user who can see the one variable you moved can direct the next iteration themselves.
If the person wants to run it on AI Content Drop, end with one link that carries
the prompt, URL-encoded with spaces as %20:
https://aicontentdrop.com/chat#prompt=<url-encoded prompt>. Say what opening
it does: a signed-out visitor is asked to sign in first; on a signed-in account
the chat on aicontentdrop.com sends the prompt at once; a chat reply costs 1 to
3 credits; and a video or image is made only when the person presses the card
that shows the model and its credit cost.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

