Nine hundred words. Camera, lens, four light sources, the grain of the wall, the melancholy of the afternoon. The take came back beautifully lit and containing the wrong person.
That is the failure long prompts produce. The model weights the opening of a brief hardest, so if the first paragraph is about light, it will build a shot about light and cast somebody to stand in it.
The short version. H3 accepts prompts up to 7,000 characters, but order decides the take: subject, action, camera, light, sound. Seeds are accepted on most tiers and rejected at 960P.
Why does order beat length?
I rewrote it as five lines, in this order: subject, action, camera, light, sound.
Two things changed immediately. The subject stopped changing between takes, and the takes got shorter to write. The 7,000-character ceiling is still there — it is for dialogue-heavy scenes, not for describing a room's melancholy.
Three rules I now follow without thinking:
- Subject before camera. Always. A perfectly framed shot of the wrong person is the most expensive kind of almost-right.
- One beat per generation. Two verbs produce a compromise between them, not two beats.
- Delete what you cannot photograph. "Cinematic" is not a setting. "35 mm, shallow depth, backlit haze" is.
How do seeds make a comparison honest?
Seeds are accepted on most tiers and rejected at 960P, which is its own small constraint.
They exist so that a comparison has one variable. Hold the seed and change a word: if the take moves, that word mattered. Hold the words and change the seed: if the take moves, the word did not matter enough to survive the roll.
Without that discipline I was comparing two things at once and calling the result intuition.
What does the Context-IR compiler change?
When a brief fails in a way I cannot attribute, the compiler rewrites it: subject, action, camera, light and audio, ordered the way the model reads.
It charges by usage, shows the rewritten version as a preview, and leaves my original untouched until I accept it. That last part matters more than it sounds — half the value is reading the rewrite and seeing what I had buried in paragraph three.
It is a rewrite, not a reroll. It cannot fix a bad idea, only bad ordering.
What no prompt can fix
Some limits are structural. 1080P stops at ten seconds. 2K starts at four. A dialogue beat routed to an action preset will move, whatever the wording. And if continuity matters and there is no reference image, the take will drift by design.
Recognising those early is the difference between a prompt problem and a workflow problem. Prompt problems are cheap to solve; workflow problems I was solving with adjectives for a week.
Prompt polish, the compiler and the tier table are on the H3 page.
I work on Callirra, so the compiler behaviour described here is the shipped one rather than a reading of a changelog. If a rewritten brief works better for you than it does for me, the before-and-after pair is what would help other people — post it and I will link it.