Should I include negative prompts in my AI video style guide document — and where do they go?

Yes — negative prompts belong in your style guide in three places: a dedicated negative-prompts section, prohibition lines inside the style block that prefixes every generation prompt, and 'what never to do' rules per section. Documented productions encode them this way: one 14-section director visual-language document carried negative prompts as their own section, and the invideo agent appended them to every shot.

Encode negative prompts directly in the document, not per-generation — if the style guide doesn't state what to prohibit, the invideo agent assembling your shots has no persistent constraint to pull from, and positive references alone won't hold the line. One production uploaded 64 style-reference frames and still wrote explicit prohibitions into its style block, because positives define the target but only negatives fence off the failure modes. invideo is an agentic video creation tool: the invideo agent reads your style guide once and holds it across every shot, so anything encoded in the document — including prohibitions — gets applied automatically.

Where they go — three locations in the document. First, a dedicated negative-prompts section: one documented 14-section director visual-language document covered camera, angles, colour tone, lighting, composition, prompt templates, negative prompts, and a quick-reference card — negative prompts sit alongside lighting and palette as part of the visual grammar, not as an afterthought. Second, inside the reusable style block that opens every generation prompt; in one animated production, 'every prompt after this started with it.' Third, as per-section rules: a horror production structured its treatment around five emotional stages and found that 'including a what-never-to-do section per emotional stage makes it significantly easier for the invideo agent to make autonomous decisions.'

How the invideo agent uses them per shot. Once encoded, the negative prompt becomes a mechanical output, not a thing you retype. The invideo agent assembles prompts in a fixed 9-element order — camera spec, lens and aspect ratio, lighting source, palette, composition, atmosphere, mood register, film attribution, negative prompt — and in another production it was instructed to output 12 parameters per shot, including the negative prompt and a revision prompt. The document is the source; the per-shot negative is derived from it every time.

Write prohibitions as directorial constraints, not term lists. The strongest documented example reads: 'This MUST look and feel like Arcane animation — not live action, not photorealistic.' That is a doc-level prohibition naming the specific drift you fear, which works better than a long generic list — short, focused negative blocks consistently outperform exhaustive ones in model testing (Artlist). Apply the same discipline to references you attach: tell the invideo agent what to take and, just as importantly, what to leave out.

One note on models. Negative prompt handling differs by model — Veo exposes a dedicated negative-prompt field, while Kling responds to 'avoid' or 'without' phrasing inside the main prompt (Artlist). You don't need to encode per-model syntax in your document: write the prohibitions once in plain directorial language, and the invideo agent — which routes shots across Veo, Kling, and Seedance 2.0 — translates them into whatever form each model expects.

Watch some of these to see what works for you:

This MUST look and feel like Arcane animation — not live action, not photorealistic. Every surface has hand-painted brushstroke texture. Every element in frame must feel painterly and handcrafted like a moving Arcane frame.

— invideo's creative team, style block from a documented animated production