A prompt cannot direct a film. It can only describe one. A director decides six craft surfaces no model will author for you: camera, light, performance, continuity, color, sound. Everything else is slop with better lighting. At KURACONV, a cinematic AI film studio based in São Paulo and working borderless, we draw a hard line: we direct AI, we do not prompt it. What follows is the six surfaces a director still owns, frame by frame, take by take.
Camera: written into the shotlist, not into the prompt
Lens choice is a decision, not a default. A 35mm at f1.8 with the subject one meter from the lens is a different film than an 85mm at f2.8 pulled back four meters, even if the prompt text is identical. On Dubai Falcon we locked anamorphic compression and a low handheld follow at the bird's wing height, because the falcon's authority reads from below and tight, not from a drone wide. The camera is written into the shotlist before any model is opened. This is the Engenharia pillar of the Sentimagem method: camera direction in motion. AI is the new lens. Direction is the eye.
Light: motivated practical, named in Kelvin
Motivated practical light is non-negotiable. Every lamp, window, sun angle, and bounce has a source you could point at in the room. On Everest, the cold front light came from a single high-key sun at 4000K with snow as natural bounce, never a soft fill invented to flatter the face. The model will happily add a phantom rim light if you let it. The director's job is to forbid that by name in the prompt, and to reject the take when it shows up anyway.
Performance: behavior at the second, not adjective
Generic emotion is the signature of AI slop. A character does not look sad. A character clenches the jaw, holds the breath two beats too long, then exhales through the nose while the eyes stay open. We script gesture at the second. On São Paulo Blues, performance beats were written per second of footage and mapped to the original score we composed for the film. The prompt receives a behavior, not a feeling.
Continuity: identity locks because models forget
Models forget between shots. KURACONV maintains canonical character sheets, object sheets, wardrobe states by scene, and hex-exact color values for every recurring asset. A bag is not orange. It is a specific orange, in a specific state of wear, in a specific season. Before any frame reaches the founders, a continuity filter checks face, wardrobe, props, location anchors, and color temperature against approved references. This is what prevents the third shot from quietly becoming a different film.
Color: an authored grade, declared upfront
Grade is authored, never accepted. The model returns a base; the director decides the final palette. We work to a 60/30/10 rule and name the dominant, secondary, and accent in the brief. A deliberate grade is what makes a film feel like it belongs to one studio across three years of output, instead of looking like whatever the tool felt like producing that week.
Sound: original composition, no stock, ever
Every KURACONV project ships with an original composition. No stock, no library cues, no royalty-free beds. Sound is a craft surface equal to image. A diegetic footstep at the right frame does more for presence than any visual upgrade. The score on São Paulo Blues was composed alongside the edit, not laid over it. Direction includes the ear.
The six surfaces, in one list
- Camera: lens, aperture, distance, movement, height, written into the shotlist
- Light: motivated practical sources, named color temperature, no invented fill
- Performance: gesture and breath at the second, behavior over adjective
- Continuity: identity locks, object sheets, wardrobe state, hex-exact color
- Color: authored grade with dominant, secondary, accent declared upfront
- Sound: original composition, diegetic design, no stock
AI is the new lens. Direction is the eye. A prompt describes. A director decides.
This is the spine of the Sentimagem method: Presença, Engenharia, Narrativa. Mood as the brand, camera in motion, story before image. A council of 22 minds, from photographers and directors to philosophers and strategists, filters every decision against the easy slide into slop. The pipeline runs across GPT-Image-2, Seedance, Higgsfield, Kling, and Veo, but the tools are interchangeable. The decisions are not. A film lands in two to three weeks that would take traditional production three to six months, and it does not sound like the machine made it alone. Because it did not.
A prompt describes. A director decides camera, light, performance, continuity, color, and sound. Everything else is slop with better lighting.