The Wayback Machine - https://web.archive.org/web/20260215170838/https://community.openai.com/t/ai-in-game-development-gamedev-tips-tools-techniques-and-gpt-llm-agent-integration/1372841
As a longtime member here (5+ years and counting!), Iām excited to kick off a general discussion thread dedicated specifically to the intersection of game development and AI-powered tools.
What Weāre Looking For:
Practical tips and techniques for leveraging AI in game development workflows (e.g., AI-driven art, procedural generation, intelligent NPC behaviors, or adaptive gameplay).
Helpful insights into integrating GPT models, AI assistants, and generative tools into various stages of gamedev.
Recommendations and experiences using specific tools, APIs, frameworks, and services.
Guidelines:
Focus primarily on sharing knowledge and actionable advice.
Keep promotional content limited! Occasional project showcases that highlight effective AI use are fine.
Support each other by answering questions, providing resources, or offering constructive feedback.
Letās build a valuable resource together!
What AI techniques or tools have significantly enhanced your gamedev projects?
I miss ChatGPT browserā¦ Hereās Claude playing AoE2, an ancient RTS (thatās still played today!)
The video got me thinking about AI for game testing not development per seā¦ Would be great to give AI your game and say āplay this 10,000 times real quickā and then have a chat about itā¦
Trying to throw out some interesting conversation starters for us as I want to see this thread go somewhere!
Yeah, I think a custom tool would do that better, especially combat stuff. Maybe an LLM or multi-modal model for playing the game slowly while the developer observes?
Iāve been looking at a lot of agents in games lately. The Among Us one upthread had a limit of 2k token PER SECONDā¦ it was a local model, but still.
yeah, thatās not that hard to do.. you can even make it without leaving the house (at least I could do it because we have partnered with a company that has planes and does orthofotos / lidar - but we do boring stuff like flood simulations for insurances instead).
Hereās a puzzle thatās been casually plugged at for months (about 4x too large to post code). Your network has been scrambled by an AI! (actually by random.choice)
āHardā rules (merely scoring) is that you must click or right-click to rotate a tile in the correct direction to align it, a single try. Shift-click will let you mark a tile you determined correct. Tile errors tracked.
The work beyond other gameplay yet seen in othersā is the Prim tree generation tuning. It can be very pipey, branchy, or leafy.
Gameplay about as long as reasoning model thinks, for your downtime. The next iteration is to make each tile object remember how particularly you messed up with over-rotation, or with better live statistics (or hiding any feedback or even tile marking to go āinfuriatingā).
GPT-x-Pro only runs on Responses API. I will not be a statistic they can include in their āover 50% of our API developers use Responsesā - because of their own model gating. Thatās a statement with as much value as ā100% of our API developers have moved off GPT-4.5ā. Plus, it is the worst for going in its own direction and diverging from the supplied task and input.
Hiccups: will be encountered using GPT-5.2, which is nice for patching when you have articulate tasks, but it is no genius. After a run of several OpenAI models was a multi-turn dead end in a task of fulfilling the need to have a frustrating-gameplay Minesweeper, (such as my one provided example of a uncertain tile configuration being the only thing 5.2 could replicate to poison a game board, and not following āneeds a solverā instructions at all), what actually delivers is a one-and-done different vendor:
(the delivered code is good though - press the Easter egg āwā key in my frivolous-vibes Solitaire)
Things Iāve shared here are basically āthe sourceā with no backend; theyād run the same if you self-hosted or even just opened the HTML, which is the point of sharing. The Netwalk game above is Python for you to run locally, where I provided a file download link for local app that gives you an experience similar to a compiled win32 desktop UI. It could be shared on a dialup BBS.
(Netwalk was originally started by Gemini Pro, fully-realized pygame-based until I got to my limit with that library without native UI controls).
One thing I have learned using GPT-style tools in gamedev is that they are most powerful when used to accelerate experimentation rather than replace playtesting. I have had much better results letting AI help me explore many small variations quickly, then using real player behavior to decide what survives. Optimizing for āleast resistanceā in tools often produces systems that are efficient to build but lifeless to play ā so I now treat friction as something to relocate into play rather than eliminate.
Iāve been using 5.2 pro in ChatGPT. Nothing fancy and I canāt give it the entire script usually, but helpful Iāve found. Might be because Iām not a dedicated developer and I tend to tinker. I know just enough to tell it what to avoid. Iāve caught it doing things just to mask the problem rather than fix it.
Iām really interested in seeing workflows from others. Iāve been pointing people to this thread outside the community, so hopefully we have more game devs show up. Takes something interesting to get people to stick around, and Iām hoping this helps us attract new devs.
Might make some time to get agents running around in my dungeon on a private instance. If I do, Iāll be sure to share with the group.
Thereās a rather well-known software developer who runs the chat that I hang out inā¦ who pretty much swears by codex now.
Toss a couple of prompts at your gpt about the differences between it and codex when coding, that might help you organize your flow to higher efficiency.
) of houses canāt be too hard) and a cesium + threejs bridge
