The author says a one-line prompt improved NPC dialogue performance in a game cutscene with GPT-6.2 Astra. The prompt focused on natural gestures and also invited the model to suggest other improvements.
Suitable tasks: Improving NPC actions and performance in game dialogue scenes; asking the model to inspect related details beyond a specified direction.
Unsuitable tasks: Assessing Astra's consistency or overall quality across other game engines, animation pipelines, or character scenarios based on this example.
Model version used: The author says they used GPT-6.2 Astra; the specific deployment or client was not stated.
Client, agent, or API used: Not stated.
Recommended reasoning effort and parameters: Not stated.
The author's original prompt:
The NPC in this sequence feels fake. Please make him do realistic gestures that you would see when people speak in real life & any other improvementsThe original post only says this was a one-shot prompt and does not provide reproducible steps. According to the author's description, the prompt produced an improved result.
The author says this was the result of one attempt to improve a game scene and that it worked very well.
The author lists these results: the NPC made less eye contact with the protagonist; because the NPC hated his job, he acted withdrawn, with gestures resembling the “five stages of grief”; and when the character turned around, the dialogue sounded quieter and more muffled to match the change in mouth position.
The author also jokingly says props and clothing in the rain scene could be more realistic, giving a lightsaber and a bikini as ironic examples.
All these details are the author's written description of the result; this article did not independently verify the visuals or audio in the video.
This is the author's account of a single use, not an independent test, and it does not prove that the improvement was caused entirely by this prompt.
The original post does not disclose the game engine, scene context, input materials, system prompt, tool permissions, model parameters, scope of changes, or any human edits.
“any other improvements” authorizes the model to expand the scope of changes on its own. For use in a real project, specify what must not be changed, asset constraints, and acceptance criteria.
On 2026-11-19, the original post was opened through Tabbit. After the page loaded, “Show original” was used to verify the English prompt and the author's description of the results. The prompt is preserved verbatim. The media in the post was not independently played or analyzed, and the task was not reproduced.
GPT-6.2 Astra