Explainer Video Prompts (Concept to Clip)

TL;DR: Explainer video prompts succeed or fail on metaphor selection, not on shot count. Pick one visual stand-in for the abstract idea, write a shot list against it, then match each shot to what the model actually controls: duration, resolution, and, on Kling, a real multi-shot field. Voiceover timing and on-screen text still get assembled afterward.

One honesty note before the pipeline. Prompt Architects writes and structures the prompt text; it does not render the clip. Nothing in this post generates a video. This is a guide to writing explainer video prompts that give whichever model you use, Veo, Kling, Runway, or another, something concrete enough to actually animate.

Why Are Abstract Concepts the Hardest Thing to Prompt for AI Video?

Because there's nothing to point the camera at. Ask a video model for "a person walking through a city at dusk" and it has a subject, an action, and a setting. Ask it for "compound interest" or "zero-trust security" or "a smoother onboarding flow" and it has none of those, because the concept isn't a physical thing. The model will still produce something, usually a person gesturing at a laptop or a vague swirl of light, but that output communicates nothing to a viewer who doesn't already understand the concept.

The fix is metaphor selection, and it's genuinely the hard part of this whole exercise. A snowball rolling downhill and gathering size stands in for compound interest. Interlocking gears stand in for a workflow that used to require manual handoffs. A locked door with a second, smaller lock behind it stands in for defense in depth. None of these metaphors are clever inventions; they're the same visual shorthand a human illustrator would reach for. What changes with an AI video model is that the metaphor has to survive being described as one continuous physical action, because that's the only vocabulary the model reliably renders well. "Compound interest grows over time" is not renderable. "A snowball starts small and gets bigger as it rolls downhill" is.

Test a candidate metaphor by writing it as a single sentence with a subject, a verb, and a visible change of state. If you can't do that in one sentence, the metaphor is still an abstraction wearing a costume, and it'll render the same way the original concept would have.

The Five-Stage Pipeline: Concept to Clip

Every explainer video prompt that actually works passes through the same five stages, whether it ends up as one generation or six:

1. Concept. The one sentence you're actually explaining, written in plain language, no metaphor yet. "New users abandon setup because there are too many steps before they see value."
2. Metaphor. The single concrete visual stand-in that survives the one-sentence test above. "A hiker dropping items from an overloaded backpack, one at a time, until the load is light enough to keep climbing."
3. Shot list. Break the metaphor into two to four beats, each with a clear visible change of state. Most working explainers don't need more than that; see the FAQ below for why more shots usually signals a concept that hasn't been narrowed enough.
4. Per-shot generation prompt. For each beat, write subject, action, camera framing, and style, then set the fields the model you're using actually exposes: duration, resolution, aspect ratio, and, where it exists, a multi-shot structure. This is the stage where vendor differences actually bite, covered below.
5. Assembly. Stitching shots together, adding voiceover, on-screen text, and music happens after generation, in an editor. No model checked for this post accepts a request field for any of those three things. That assembly pass is also where the two most common follow-up questions live: how the cut between shots actually reads, covered in transitions in AI video prompts, and how to direct any music underneath the clip, covered in music direction in video prompts.

What Actually Generates Inside One Shot, and What Do You Assemble Afterward?

This is the split that most explainer-prompt advice skips, and it's the one that actually determines whether your plan is realistic.

Genuinely generation-time, confirmed on the vendor's own parameter tables: duration, resolution, and aspect ratio. On Veo 3.1, durationSeconds accepts "4", "6", "8", and Google's own parameter table pairs it with this exact constraint: Must be "8" when using extension, reference images or with 1080p and 4k resolutions. resolution accepts "720p" (default), "1080p", and "4k", though the Lite variant doesn't offer 4k at all. aspectRatio is "16:9" (default) or "9:16" only, on every Veo 3.1 variant checked, so an explainer shot in a square or ultra-wide frame simply isn't an option on this model. See our full AI video duration parameters by model reference for every vendor's exact enum, since the numbers above are Veo's alone.

Genuinely generation-time, but written as vocabulary rather than set as a field: pacing. Google's own video generation prompt guide has a section titled "Temporal elements" with a "Pacing" sub-heading, giving "slow-motion", "fast-paced action", and "time-lapse" as example phrases. That's a real, vendor-documented way to control pacing, and it belongs inside your shot description, not in a settings object, because no vendor checked for this post exposes a numeric pacing parameter.

Not generation-time on any surface checked: voiceover timing, on-screen text placement, and the pacing between separate shots when your model has no multi-shot field. None of these are request parameters on Veo, Sora, Kling, or Runway's core generation endpoints. They happen in an editor, after the clips exist. No tool referenced in this post auto-edits a rough cut into a finished explainer or auto-publishes anything; you're still assembling the pieces by hand, or with separate editing software, once generation is done.

Getting this split right before you write a single prompt saves you from planning a workflow around a control that doesn't exist.

How Do Kling, Runway, Veo and Sora Handle Multi-Shot Sequences?

Differently enough that it changes how many separate generation calls your pipeline needs.

Kling 3.0 Omni has a real request field for this: settings.multi_shot, a boolean that defaults to true. When it's on, the prompt text itself carries the shot structure, using a documented grammar: "shot n, m, words;", where n is the shot's sequence number (1 to 6 shots supported), m is that shot's duration in seconds, and words is the shot's own prompt text, capped at 512 characters. The per-shot durations have to sum to the overall settings.duration, which itself accepts any integer from 3 to 15.

Runway ships a dedicated recipe for this rather than a field on its core generation endpoint: POST /v1/recipes/multi_shot_video. In custom mode it takes a typed shots array of three to five entries, each with its own prompt (3 to 512 characters) and duration, and the Recipe assembles them into one output. Auto mode takes a single story prompt and lets the Recipe decompose it into shots for you.

Veo 3.1 and Sora 2 have no equivalent field. Veo's documented request parameters are aspectRatio, durationSeconds, personGeneration, and resolution, full stop, there's no shot-count or shot-list parameter anywhere in that table. Sora's request shape is similarly single-shot. Practically, that means an explainer built on either model is a series of separate generation calls, one per beat, stitched together afterward rather than requested as one multi-shot job.

Where Do Voiceover and On-Screen Text Fit?

Nowhere inside the generation call, on any of the four surfaces above. Once your shots exist, whether that's one Kling multi-shot generation or four separate Veo calls, voiceover and on-screen text are an editing pass, not a prompt. Write the voiceover script against the total runtime you've already committed to in the shot list, and budget roughly 2.5 words per second of clip at a natural pace, tighter if the read is meant to sound like an ad.

If your project genuinely needs the word-count-against-duration math worked out in detail, marking beats and timing lines against seconds, that's a separate, deeper problem than this post covers, and it's worth its own pass rather than a paragraph here. Treat script-to-clip pairing as the next step after this pipeline, not a step inside it.

A Worked Example: Prompting a Compound-Interest Explainer, Shot by Shot

Concept: "Money left to grow compounds faster the longer you leave it alone." Metaphor: a snowball rolling downhill, gathering size as it goes. Three shots: small and slow, then large and fast, then landing on the outcome.

On a model with a real multi-shot field, that's one Kling generation:

prompt (settings.multi_shot: true, settings.duration: 12):
"shot 1, 3, close-up on a single snowball starting to roll at the top of a
snow-covered hill, soft morning light; shot 2, 5, wide shot as the snowball
has grown to twice its size and gathers speed down the slope, motion blur
in the trailing snow spray; shot 3, 4, the now-massive snowball flattens a
small snow fort marked with a dollar-sign flag, wide aftermath shot as dust
settles;"

On a model with no multi-shot field, the same three beats become three separate calls:

Shot 1 — config: { aspectRatio: "16:9", durationSeconds: "4", resolution: "720p" }
prompt: "Close-up, slow motion: a single snowball starts rolling at the top
of a snow-covered hill, soft morning light."

Shot 2 — config: { aspectRatio: "16:9", durationSeconds: "6", resolution: "720p" }
prompt: "Wide shot, fast-paced action: the snowball has grown to twice its
size, thundering down the slope, motion blur in the snow spray."

Shot 3 — config: { aspectRatio: "16:9", durationSeconds: "4", resolution: "720p" }
prompt: "Wide aftermath shot: the now-massive snowball crashes into and
flattens a small snow fort marked with a dollar-sign flag."

Same metaphor, same three beats, two very different numbers of API calls, because the field support underneath is genuinely different.

Should You Still Prompt Sora 2 for a New Explainer Project?

No, not for anything you plan to keep running past this month. The distinction that matters is which Sora surface you mean: the consumer web and app experiences are already gone, discontinued April 26, 2026, while the API that sora-2 and sora-2-pro run on is still live today but scheduled off on September 24, 2026, with OpenAI's own deprecations table listing no replacement model for any of the six retiring entries. If you have existing Sora prompts worth keeping, treat this as a migration deadline, not a someday task, and move the underlying concept and shot list over to Veo or Kling rather than the exact prompt text, since none of the field names carry over directly. Our Sora 2 migration guide covers that field-by-field, model-by-model.

Crowds, Presenters, and Background Characters in Explainer Shots

Most explainer videos need at least one human figure, a presenter gesturing at the metaphor, a customer reacting to the outcome, sometimes a small background crowd in an office or classroom scene. Keeping a single presenter recognisably the same across separate shots is its own problem, covered in our guide to video character consistency. If your shot list calls for a filled room or a background crowd rather than one named character, that's a different and arguably harder problem: crowd and background character prompts covers why those figures are where most models produce their worst artefacts, and what to write instead of hoping for the best.