Using GPT-4 for Game Text Review: A Prompt‑Engineering Case Study

GPT‑4, OpenAI's large language model, can assist in writing, chatbots, programming, news generation, and game design. The article explores its use for checking massive, frequently updated game text, a task that traditionally suffers from low inspection frequency and high risk of inappropriate content.

The first attempt with a vague prompt yielded partial results: fewer output cases than input, missing clear conclusions for some items (e.g., "Apple Inc."), incorrect judgments (e.g., "盗窃星晶宝盒"), and extraneous explanations, prompting the need for prompt improvement.

Prompt debugging proceeded in three steps: (1) provide a few concrete examples, (2) specify a strict output format, and (3) ask the model why it made certain decisions. Additional safety criteria were added, such as flagging illegal, immoral, or insulting terms, and filtering celebrity names.

After these refinements, GPT‑4 produced more consistent conclusions for each term, reduced unnecessary explanations, and better adhered to the desired output format, though some edge cases still required manual review.

The final production prompt instructs the model to act as a text editor for teenage‑targeted games, evaluate Chinese phrases for inappropriate moral guidance, tolerate mild profanity, violence, and purchase terms, but reject any phrase containing celebrity names or encouraging illegal behavior.

A comparison with GPT‑3.5 shows GPT‑4’s larger scale and improved understanding lead to higher accuracy and consistency, making it the preferred choice for large‑scale game text moderation.

The article concludes with a recommended iterative workflow: define a clear prompt, analyze model feedback, adjust objectives and prompts, and repeat, highlighting the value of prompt engineering in leveraging AI for game development tasks.