How to Generate a Prompt From a Video: The Eight-Slot Template

Most people write video prompts as sentences, and most video prompts fail as sentences. A generator does not parse grammar; it pattern-matches on caption-shaped fragments. So the reliable move is to stop writing and start filling in slots — a fixed eight-field skeleton that you complete from the clip in front of you. Here is the template we use, how an extraction tool populates it, and the iteration loop that turns an almost-right render into the one you keep.

Why templates beat prose

Training captions for text-to-video models are dense, comma-separated and noun-heavy: “a wide shot of X, golden hour backlight, slow dolly in, shallow depth of field”. When you write flowing prose — “I want something that feels like a memory of a beach” — you are asking the model to infer a caption from a mood, which it will do inconsistently. A template keeps you on the distribution the model was trained on, and makes every run comparable: same slots, changed values, visible cause.

The eight slots

1. Subject — the thing in frame, one noun phrase. “A ceramic coffee cup”, not “a beautiful morning moment”.
2. Action — one verb phrase. “steam curling upward”. If nothing moves, say so: “locked-off, no subject movement”.
3. Lens and format — focal feel and capture medium: 35mm, macro, anamorphic, phone vertical, 16mm grain.
4. Camera move — exactly one verb: dolly in, whip pan, crane up, handheld drift, static.
5. Light — source, direction, quality: “window key from camera-left, soft, low contrast”.
6. Grade — palette and contrast, as colour words: “warm highlights, teal shadows, lifted blacks, fine grain”.
7. Pacing — duration and cut rhythm: “8 seconds, single take” or “three quick cuts on the beat”.
8. Mood — last, and only one or two words: “uncanny”, “cozy”. Mood carries the least weight; spend your character budget earlier.

Assemble in that order, comma-separated, and keep it under about 60 words. Slots 3–6 are where an extracted prompt earns its keep, because they are the ones you cannot see while you are watching the thing.

Filling the slots from an extraction

Read the clip yourself, then let SocialToPrompt fill the gaps, then reconcile. Paste the link (or upload MP4, MOV or WebM) and the engine returns the fields it inspected for — motion, camera work, composition, style, lighting and timing — after a run the site says typically takes 20–40 seconds. Drop its values into the eight slots, and mark anything you disagree with in brackets; that bracket list is your next iteration queue.

Two slots resist automation and you should own them. Lens and format depends on where your clip is going (a 9:16 phone ad does not want anamorphic bokeh), and mood depends on intent the clip cannot tell you. Everything else is fair game for extraction — see how to extract a prompt online for the mechanics.

The one-variable iteration loop

Change one slot per run, and write the changed value in the filename. The loop that converges:

Run	Change	What you are testing
1	Baseline from template	Did the subject and framing land?
2	Camera verb only	Is the movement word the model knows?
3	Light, direction or quality	Does the mood come from lighting rather than colour?
4	Pacing / duration	Is the render falling apart because there is too much to do?
5	Grade as explicit palette	Is the look a palette problem rather than a subject problem?

Five runs with five single changes will teach you more than twenty rewrites, and it keeps the cost predictable — on a credit-metered tool that matters, which is why we do the arithmetic in what the free tier buys.

Versioning your prompts

Download the .txt the tool offers, then rename it into a sortable pattern: 2026-10-13_coffee-turntable_v3-dollyin.txt. At the top of the file, keep three lines of context the prompt itself cannot carry: the source link or file, the target model, and what you changed from the last version. Two months later that header is the difference between reusing a good prompt and reverse-engineering your own work.

A library of these files also exposes patterns: which camera verbs your model actually obeys, which grade words survive across subjects, which pacing length your audience watches to the end. That is the education hiding inside the tool.

Common slot mistakes

• Two cameras in one slot. “Dolly in then orbit” in four seconds renders as mush; pick one.
• Adjective stacking. “Stunning, breathtaking, epic cinematic” adds weight to nothing. Concrete beats enthusiastic.
• Describing the reference rather than the recipe. “Like the Nike ad” is a request the model cannot fill; describe the mechanism instead.
• Negations. “No people” often invites people. Write the positive: “empty street”.
• Ignoring aspect ratio. A slot-3 problem masquerading as a composition problem — vertical clips need vertical framing words.

For the methodology behind choosing which clips to template in the first place, see video prompt reverse engineering, and for the fields themselves, converting a video to an AI prompt.

Frequently asked questions

What template should I use to write a prompt from a video?

An eight-slot skeleton in this order: subject, action, lens and format, camera move, light, grade, pacing, mood — comma-separated, under about 60 words. Early slots carry more weight with the renderer, which is why mood goes last and camera detail goes third or fourth rather than being buried in prose.

Can a tool generate the prompt for me?

Yes — an extractor reads the clip and returns most of the fields, typically including camera behaviour, lighting, pacing and style. You still own two slots yourself: lens and format, which depends on where the clip is going, and mood, which is intent rather than observation.

How do I stop changing five things at once when I iterate?

Change one slot per run and put the changed value in the filename. Five single-variable runs answer five questions; twenty rewrites answer none, because you cannot tell which edit helped.

Why do adjectives and negations hurt my renders?

Stacked intensifiers carry no visual information, so they dilute the concrete words around them. Negations are worse: models frequently render the thing you told them to exclude, because the token is present. Write the positive state instead — 'empty street', not 'no people'.