How the agent knows each model's prompting
When you generate through Versely's chat, the agent isn't guessing what a model wants — and it isn't “trained on every model” either. Its model knowledge is four concrete mechanisms:
get_model_input_schema— a tool the agent calls before generating that returns the model's exact input fields, required fields, allowed values, defaults, and min/max bounds. The parameter tables on these pages are that same surface.- 12 family rules in the prompt enhancer — when it rewrites a rough prompt, per-family rules fire by model name (Midjourney wants comma-separated descriptors, Kling wants named camera movement, Wan wants plain description plus style keywords, and so on). Each page quotes its family's rule verbatim.
- A per-provider speech guide — for TTS, the agent knows which providers take inline emotion tags (and each one's syntax) and which take emotion as an API parameter where inline tags would be read aloud.
expand_movie_scene— in movie flows, brief scene ideas are expanded into detailed cinematic descriptions before generation.
These pages publish that grounded knowledge so you can use it by hand — and anything on them that is general craft rather than a documented model fact is labelled as such.
t2v · 13 guides
Text-to-video: scene narrative, camera movement, temporal flow; negative prompts where supported
talking video · 9 guides
Lipsync/avatar/audio-driven: text or audio input, delivery and framing controls
i2v · 8 guides
Image-to-video: animating a start frame; motion description relative to the input image; UGC hook pool applies
t2i · 6 guides
Text-to-image: composition, lighting, style; family tips (Midjourney/Flux/Imagen/etc.) drive syntax
reference to video · 5 guides
Reference-images-to-video: multi-image identity/style consistency
extend retake · 3 guides
Extend/retake: continuing or re-rolling an existing generated clip
video edit · 3 guides
Video-to-video editing/restyle/segmentation: instruction-style prompts over an input video
utility enhance · 2 guides
Upscale/enhance/background-removal: mostly parameter-driven, little to no prompt
motion control · 2 guides
Motion transfer: reference video + character image; prompt describes retained vs replaced elements
first last frame · 2 guides
First/last-frame and transition: describing the in-between motion
storyboard · 1 guide
Multi-shot storyboard: per-shot scene descriptions (Sora 2 Pro Storyboard shots schema)
Looking for ready-made UGC hook prompts rather than per-model syntax? That's the prompt library — a browsable pool of proven hook prompts, where these pages are references for how each model wants any prompt written.
Frequently asked questions
How does the Versely agent know each model's prompting rules?+
Four real mechanisms: it fetches each model's exact input schema with the get_model_input_schema tool before generating; its prompt enhancer applies 12 per-family rewrite rules (Midjourney, Flux, Imagen, Kling, VEO, Wan, Seedance, Sora and others); its speech guide encodes each TTS provider's tag scheme; and expand_movie_scene rewrites brief scene ideas into cinematic descriptions in movie flows. These pages publish that same grounded surface.
Are these guides written from the models' actual schemas?+
Yes. Every parameter table is the model's real input schema — the same one the agent reads — and every quoted rule is verbatim from the product. General craft advice is labelled as general technique, not presented as a model fact.
Why do some models share one prompting page?+
When two variants of a model expose an identical prompting-relevant input surface (same params, enums, and family rules), they share one canonical page and the sibling URL redirects to it — one honest reference instead of near-duplicate twins.
Or skip the syntax entirely
The all-in-one AI studio for creators. 60+ models for video, image, voice, music and lipsync in a single app.