What stopped working
Model-specific magic strings ('cinematic, 8k, hyperrealistic, unreal engine') are actively harmful now — they push toward a generic Midjourney-2023 aesthetic that modern models were trained away from. Weight syntax like (word:1.4) is ignored by every closed model. Negative prompts still work in Flux but do nothing in Sora / Veo / Kling.
What works: the shot-brief format
The most reliable 2026 prompt is a shot brief a DoP could hand a camera operator. Six slots:
- Subject — who or what, one sentence
- Action — what they do across the clip
- Camera — angle, movement, lens feel
- Lighting — direction, colour, quality
- Setting — location and time
- Mood or reference — one word or one film reference
Concrete example
Weak: 'A woman walking in a city, cinematic, 8k.'
Strong: 'A woman in a black wool coat walks left-to-right down a Tokyo alley at dusk. Handheld, following, 35mm feel. Neon rim-light from a red sign camera-right; overcast sky. Light rain on asphalt. Mood: Wong Kar-wai.'
The strong prompt hits a keeper on the first roll in most models. The weak one takes four to six.
Prompt length: 25-60 words is the sweet spot
Under 25, the model improvises too much. Over 60, it starts dropping details. Every closed model in 2026 has a soft ceiling around 70-80 words before quality drops.
Practice in the studio
Put this into practice in the studio — under a minute to your first result.
Practice in the studio →