The one video model with a real negative field — around 2,500 characters, and capacity is not permission.
- A real, separate negative prompt field — roughly 2,500 characters, alongside 2,500 for the prompt.
- Do not write "no". The field already treats entries as exclusions.
- Short beats long. Around five focused terms, aimed at glitches you have actually seen.
- Released 7 February 2026 — 3.0 added native audio and longer durations.
- Stable Diffusion habits transfer here more cleanly than to any other video model.
Field behaviour and limits are vendor-documented; the tuning advice is established practice.
The formula
Template:
PROMPT: [subject] [action] in [setting], [camera behaviour], [lighting and style].
NEGATIVE FIELD: [term, term, term, term, term]
Worked example:
PROMPT: A ceramicist centres clay on a wheel in a north-facing studio at dusk, slow push-in, window light only, documentary realism.
NEGATIVE FIELD: blurry face, extra limbs, jittery motion, text overlay, watermark
Breakdown — what each part does
Prompt field
Subject, action, setting, camera, style — conventional structure, up to about 2,500 characters.
Negative field
Separate box. Plain terms, comma-separated. It is read as exclusions automatically.
Physics
Kling's documented strength is physically plausible motion — water, fire, fabric, impact. Prompt for it explicitly when it matters.
Does it take a negative prompt?
Plain terms only. Do not prefix with "no" — the field already means exclude. Target artifacts you have actually observed: blurry faces, extra limbs, jittery motion, text overlays, watermarks. Around five terms is the reported sweet spot; the character limit is capacity, not an invitation.
What not to do
Kling's API documentation covers the fields and limits; the practical guidance here (five-term sweet spot, targeting observed artifacts) comes from published prompting guides rather than a vendor style manual. Field behaviour and limits are documented; the tuning advice is practice.
The checklist
Before you send it:
Kling API documentation and published prompting guides. Model release: Kuaishou, 7 Feb 2026. Checked 25 Aug 2026.
Model versions and vendor documentation both move. Re-read the source before relying on a specific number. Errors are logged at corrections.
Every other sheet in this set: all model cheat sheets.