Prompting guides are per model, not per vibe
/prompting and /models/{slug}/prompting exist because Veo timestamp blocks are not Kling Omni grammar.
A mega-prompt that worked once — "cinematic, 35mm, golden hour, slow push-in, natural skin, no morphing" — is not a studio asset. It is a sentence that happened to please one model on one day. Paste it into Veo 3.1 and you skipped the timestamp blocks Google actually documents. Paste it into Kling O3 and you skipped the camera-and-collision grammar that model honors. Paste it into Seedance R2V and you skipped the references, which were the job.
Stop copying one mega-prompt across the catalog. Prompting is the hub. Each published model that has a real input surface also has /models/{slug}/prompting — parameters, enums, templates, and the rules the agent already applies. Veo timestamp blocks are not Kling Omni grammar. That is the whole reason those pages exist.
One catalog, many input schemas
Versely's agent does not "know every model" in the mystic sense. Before a non-trivial generate it can fetch the model's actual input schema: required fields, allowed values, defaults, bounds, one entry per provider route. The prompting pages publish that same surface. If a field is not on the page, it is not a lever you have.
Vibe language is almost never a field. "Make it more premium" is not duration, aspect_ratio, generate_audio, first_frame_url, or a timed block. Models that accept a prompt still do not share a prompt dialect:
- Veo 3.1 wants shot language and, for multi-beat clips, timed blocks (
0–2s / 2–5s / 5–8s), plus an explicit instruction about what we hear, because audio is native. The essay is VEO 3.1 prompting; the schema is/models/veo-3-1/prompting. - Kling O3 / Omni wants cause-and-effect motion and camera mechanics. Orbit, collision, what the camera does in response. Mood is downstream. Kling O3 prompting and
/models/kling-3-turbo-text-to-video/promptingare the t2v door; reference-to-video is a different endpoint with a different page. - Seedance 2.5 wants references more than adjectives when identity has to hold.
@files in one generate and the dedicated R2V endpoint are not interchangeable with a prose paragraph. - TTS families each have their own tag scheme. ElevenLabs tags are not a Qwen SSML paste. The speech pages say which.
Copying a Veo block into Kling is not "being consistent." It is sending a no-op paragraph and hoping the model infers a camera move you did not specify in its grammar.
What a per-model page is for
Use it when you are about to spend video credits on a named model. It will tell you:
| Look for | Why |
|---|---|
| Required fields | Missing image_url on I2V is a failed job, not a creative miss |
| Enums (duration, resolution, aspect) | "Cinematic 4K" is not an enum if the model does not expose it |
| Audio on/off | A silent model will ignore a dialogue novel in the prompt |
| Mode (t2v / i2v / R2V / first-last / extend) | Each mode is a different page on purpose |
| Family mistakes | The ones that burn credits for this family, not "AI video" in general |
Use a vibe library when you need a register — bathroom-mirror talking head, POV hands — and then translate that register into the model's fields. Hook prompts are registers. Prompting guides are schemas. Do not collapse them.
The agent already rewrites some families (Midjourney, Flux, Imagen, Kling, Veo, Wan, Seedance, and others) through a prompt enhancer. That is not a licence to feed it a universal poem. It is a reason to start from the schema so the rewrite has the right bones.
The mega-prompt failure modes
The portable masterpiece. A paragraph that includes camera, wardrobe, brand hex, product claims, negative prompt, and a joke. One model will use three of those clauses. The rest is noise that sometimes hurts. Split: stills carry wardrobe and product; the video prompt carries motion and sound; the schema carries duration and ratio.
The cross-model A/B. Same prompt, two models, "which is better?" You did not A/B the models. You A/B'd which dialect got ignored. If you must compare, write each arm in that model's grammar, same stills, same duration, same job.
The film-shaped paste. A scene list belongs in Movie Mode or a workflow, not in a single generate. Timed Veo blocks are not Sora Storyboard. Lock stills, then I2V.
Practical rule
Before you generate: open /models/{the-slug}/prompting. If you cannot name the slug, you cannot name the dialect, and you should not paste last week's paragraph. If two variants share a page, they share an input surface — use that page once, do not maintain twins.
The catalog at models is the index. The hub at prompting is the map. Vibe is a taste. Schema is a contract. Only one of those ships.
FAQ
Can I keep a studio-wide prompt suffix for brand consistency?
Yes — palette, forbidden claims, logo rules. Put those in the brand kit and in stills. Do not put Veo timestamp syntax or Kling camera grammar in a suffix that every model will swallow. Shared suffix, per-model skeleton.
Why do some models share one prompting page?
When two variants expose the same prompting-relevant inputs, one canonical page is the honest reference and the sibling URL redirects. Tier and price still live on each model page. Do not split a guide that would duplicate the schema.
Does the agent read these pages?
It reads the schema tool and family rewrite rules those pages are built from. The pages exist so you can see the same surface. If you name a model and ignore its fields, you are fighting both the model and the agent.
Where do I put a multi-shot Veo idea if not in a mega-prompt?
In timed blocks on a Veo clip, or in locked stills plus I2V, or in Movie Mode if the deliverable is a film. Three different objects. The prompting page will not turn a paragraph into a board.