I tested Seedance 2.5 the way most people probably will:

Source: @leelinai123. Model label in the archive: seedance 2.5.

The full prompt

I tested Seedance 2.5 the way most people probably will:
describe the idea in plain language → have Codex GPT-5.6-SOL (Ultra) turn it into a detailed prompt → add 4 reference images → run one generation.
This is what came back.
At first glance, it works. Same two characters. Same rainy street. Consistent left-right staging. Full 30.08 seconds at 720p, with stereo audio.
Then I checked it against the prompt, beat by beat.
The opening walk becomes a pose. A retreating parry becomes a static block. The ground slam turns into another pose. A low, continuous dash becomes a slow-motion close-up. The final framing hides the axe and foot positions, and the lightning cue lands late.
Seedance 2.5 can do much better than this. And this Codex draft isn’t how I’d structure a production prompt. That’s the point: this is a baseline, not a ceiling.
It shows what happens when a plain-language request gets expanded into something that looks complete on paper.
Long doesn’t mean professional. Detailed doesn’t mean executable.
A production-ready prompt makes the hierarchy obvious. The model needs to know which action matters most, where contact happens, what the camera has to capture, and exactly how the shot should end. If every constraint is equally important, the model ends up choosing for you.
Codex wrote the original prompt in Chinese. Reply 1 has the full English translation in standard production terminology, so you can compare the brief with the clip.
I’ll keep posting prompt/output pairs and breaking down what actually works.

Follow the original creator for context, references and their example.