Everyday Unity

How Claude Code Uses Prompt Caching I believe it's important to know what prompt caching is and how it helps save costs. On this page, I would pay attention to what resets the cache, how agents affect it, how to check your cache usage, and TTL. One advice I didn't follow before:

Compaction works in your favor when the context you discard is content you no longer need. To choose when its overhead happens, run /compact at a natural break in your work, such as between tasks, instead of waiting for auto-compaction to trigger mid-task. If you’ve gone down a path you want to abandon entirely, /rewind to an earlier turn instead. Rewinding truncates back to a prefix that is already cached, rather than building a new one as compaction does.

https://code.claude.com/docs/en/prompt-caching #ai

Meet the Unity CLI: manage Unity from your terminal I had multiple posts about how I used godot because of great cli tools and now its time to test if Unity can match it. Seems like it doesn't have a lot of stuff yet, but it allows agents to eval any code without full recompilation and domain reload. So in theory agents can do any custom work with this command. Has anyone tested it already? https://unity.com/blog/meet-the-unity-cli #cli

I transitioned my roguelike game from 2D to 3D to test Sol Ultra when it was released 1.5 weeks ago. I def can say it was a successful experiment. The game was working, but lacked a few animations I had before. I initially used original sprites as billboards in 3D world, but then eventually swapped it with very simple models. The game transformed from a simple 2d board to a stylized 3d after I threw in a few shaders and post effects. And of course the game again became slower to load. So I ran the same profiling with tracy and got the faster loading in just 5 mins. Cannot recommend it more.

Just now Open AI temporarily removed the 5h limit. It's an interesting addition to the morning discussion we had here about hitting the 5h limit with sol ultra in 18-40 mins. I hit the 5h limit twice today and have only 32% of a weekly limit left. Thanks Open AI, now I can spend my weekly limit in around 1.5 hours 💪💀

I have been testing 5.6 Sol lately and haven't seen a really big difference to the old model when I used it with medium, high, extra high effort. The only thing that surprised me was a co-op bot for my roguelike game. You press a bot run button on the host and it runs kinda optimal strategy and plays the game from all clients. It was done in around 2 prompts and works very well. I was able to catch around 6 bugs already with it. Then I added auto mode for LLM, it launches multiple clients each with their own set of launch args. And then LLM reads the console every 5 mins for each client to find anomalies or bugs. It also found a few bugs which led to an incorrect loss. Apart from that other tasks were the same in terms of speed and quality as 5.5. This morning I wanted to test 5.6 ultra and compare it against Fable 5, which I used at work to create a pretty big feature on which it worked for 2 days straight with pauses when it was hitting the 5 hour limit of $200 subscription. So I selected ultra effort, started to write a prompt and then I saw the new update available. It must make things better, right? I hit it, codex restarted, and the ultra mode is now gone 💀 Do you still have ultra effort on the latest codex version? And have you tested it against Fable? Which one is better for you? Drop 💯 if you want to hear more about my Fable experience and how it completed the feature in 2 days on its own.

12-Factor Agents for development workflows It is mostly a guide, not a library you install and build your game on top of. There are some supporting packages and workshops in the repo, but the core value is the principles for building reliable agents. The guide pushes against the "big prompt + tools + loop until done" approach. I have been using AI coding workflows for multiple years already and that's the part where things usually break on big project. It might work for a demo or prototype. But in real Unity projects with long import times, slow build pipelines, and old god singletons no one wants to touch, suddenly LLM even with agents needs logs, state, approvals, retries, and boring deterministic scripts and tools instead of a magic AI button. Few points I would steal for my game dev agents: Own the prompts. If an agent works on my project, I don't want the important prompt hidden inside a framework. I want it in the repo near requirements, diagrams, and maybe even next to scripts it can call. Same reason I like PlantUML: it can be reviewed, changed, reused. Tools are just structured outputs. "Run tests" should call a script. "Build WebGL" should call a script. "Check Addressables" should call a script. The LLM can choose what to try next, but it should not invent whether the build succeeded. Deterministic code should return the result. Pause when action is risky. An agent can prepare a PR, but I don't want it to push to master or upload a public build without explicit approval. Same for deleting assets, migrations, changing live configs, save compatibility, etc. It might sound obvious, but a lot of agent demos skip exactly this boring part. Keep the state inspectable. When an agent fails and retries, I want to see why. Unity has enough "works on my machine" mystery already. If the agent changed a prefab, fixed one error, then broke another platform, I need a history that can be inspected later. Make agents small. One agent for "prepare playable build" is okay. One for "investigate this crash report" is okay. A single agent that can modify code, assets, configs, CI, store metadata, and production data is a nice way to lose control. However I am wondering how many tokens can be wasted if there are too many small agents for each small step. Something I would like to test. Of course, the repo is written mostly from SaaS/product experience, not Unity production. But I think the principles fit game dev really well, because our pipelines are slow, stateful, full of tools, and full of things that should not be changed by "vibes". I am using my agent team at work daily and already modified it a few times based on the experience. Today I updated it according to 12-factor-agents principles, lets see if it yields better results or just turns into a bigger token drain with no significant benefits. Have you already built any agent around your game project workflow? Not just coding in Cursor/Claude, but something that runs tests, checks logs, prepares builds, or validates assets? My unity agents: https://github.com/AlexMerzlikin/unity-agent-team 12 factor agents: https://github.com/humanlayer/12-factor-agents #agent

Since May I have been working on a roguelike prototype, first in Three.js and now in Godot. I started with Three.js because it was the fastest way to check the idea. After a small playtest with my friends, when I verified that it is fun at least for me in coop, I decided to move it to a real engine. I picked Godot, even though it is my first game with it. And then I saw a long loading time. Of course, the first temptation is to ask an LLM to "optimize the game" and hope for magic. But the better loop is still the old reliable one: profile first, then optimize if needed. I checked the docs and found that Godot can be connected to Tracy. So I fed the docs to Codex and in one prompt it: - downloaded Tracy - set it up for the project - made a capture - analyzed the result - proposed fixes Then I asked it to apply the fixes, run a review with my functionality reviewer agent, and profile with Tracy again. The biggest visible win was exiting the game to the lobby. Before it took around 3-5 seconds. Now it is basically immediate. The complete result is in the pic This is not a "LLM solved performance" story. More like: LLM made the profiling loop cheaper and faster. Have you already used LLMs together with real profilers, not just for random code cleanup? Give 🔥 if you want to hear what it was like for me to port the game from Three.js to Godot. #performance #profiling #godot #tracy

Antigravity 2.0 update after a small real test. I finished 2 small tasks on my relatively small prototype with it. Nothing huge: 2 PRs, around +250 and +50 lines of code, with code review by 4 agents. Gemini 3.5 high. The result was useful, but the cost in limits was not that inspiring: around 80% of the 5h limit used on the $20 Google AI Pro subscription. It feels close to the painful Claude limits problem. And ngl, Antigravity limits were already very unclear and a sad experience for me in the past. First it looked like only 5h limits, then after reaching it a few times it changed to weekly limits. What is more, those weekly limits were updating to 1 more week when I was almost near the weekly reset, even without using Antigravity. Happened multiple times in a row until I stopped even checking it. Of course, this is not a benchmark. Just one small evening with a small prototype. But for game dev prototypes, predictable limits matter a lot. If the tool stops exactly when you found the flow, it becomes hard to trust it as a real part of the workflow. Will keep using it a bit and update whether this part is fixed or not. Have you tried Antigravity 2.0 already? Are the limits clear for you?

Find 5 differences

I switched to Codex a while ago because Claude limits were brutal for my workflow. Not even coding at first. I was hitting the 5h limit after 4-5 prompts about my GDD. For a small prototype. Opus was basically no-go for me, and Sonnet was useful, but not producing the same level of results for the work I needed. Then Codex mobile landed, and it closed the only big gap in my workflow. I could continue the same "write from phone, let the agent work on desktop" loop I had before, but inside Codex. My current experience: $20 subscription was already enough for steady progress. I finished more tasks than before, especially smaller prototype tasks and planning cleanups. For active weekend development I moved to the $100 plan. So far I am not even close to limits. I don't know if this would survive full-time work, but for weekend projects it is more than enough for me. Claude might be better now, but I haven't re-tested it yet. I heard Anthropic increased limits on their subs, so take my comparison with a grain of salt. My old frustration may already be outdated. Antigravity is the interesting competitor now. I tried the new Antigravity 2.0 after Google I/O, and the shape is very familiar - basically Codex-like agent work, but without the mobile access I rely on. I need to test it more, because Gemini 3.5 Flash might be strong enough to make the tradeoff less obvious. For game dev specifically, mobile access became much more important than I expected. It lets me keep momentum when I have a small pocket of time, and for hobby projects momentum is often the bottleneck, not raw model benchmark score. What do you use nowadays?

A follow-up on my /grill-me post from last week. After using it for about 7 days, a few observations. It asks a lot more questions than /brainstorming. Like, significantly more. I had one session with 60 questions, and I was answering them for 1.5 hours. The original task was pretty small initially. But the result? Complete set of requirements, implementation came out bug-free, following the existing project architecture. So it was time-consuming but very effective. After a few more tasks like this I changed the skill to "self-grill". Now the AI asks all those questions itself and gives me a decision list in seconds. I just scan the decisions, adjust what looks wrong, and proceed. Of course the effectiveness is lower compared to going through the full /grill-me flow properly. But for most tasks it is a much better trade-off from a time-result balance perspective. If I had to frame it: /grill-me is worth it when scope is unclear and getting it wrong is expensive. Self-grill is for everything else. Have you tried building requirements this way before coding? I am curious how others are managing the thoroughness-vs-speed trade-off in their AI workflows. #ai

"Software Fundamentals Matter More Than Ever" — Matt Pocock I recently watched a short conference talk about how software fundamentals are even more important now in the age of AI. It focused heavily on one of my favorite books, "A Philosophy of Software Design." Of course, not all the advice there applies directly to game development, but I still highly recommend checking it out and using the other concepts proposed by the author. I am already trying a /grill-me skill, and it feels like /brainstorming from superpowers. But it asks a lot more questions and helps you shape much better requirements. Anyway too early to say which one is better. What do you think about it? https://www.youtube.com/watch?v=v4F1gFy-hqg #ai

A follow-up on my Unity agent team repo. In the comments on the last post Artem asked about the custom MCP tools the tester agent uses. Here is the implementation from my prototype, sitting on top of the Coplay Unity MCP: custom actions to interact with the game, screenshots and scene hierarchy queries, so the agent can actually verify its own "manual" tests. Nothing fancy, but it closes the loop end-to-end. Feel free to adjust it for your own game logic. Fair warning: this drains tokens fast. Automated tests are of course way faster and cheaper to run, and if you can cover a case with a unit or integration test you probably should. The difference with this setup is that the agent plays your actual game against "manual" test cases, not a test harness with mocks and stubs. And on my prototype I saw real results: the agent walked through real gameplay logic, caught actual issues, and reported back in plain English. Not a replacement for the automated suite, more of an extra pass on top. Your mileage may vary. https://gist.github.com/AlexMerzlikin/a810877fb26c2325536213295ddb84c3 What is your setup for letting an agent verify its own changes in a running Unity build? #ai #agent #unity #mcp

A while ago I shared my Unity agent team repo here, and I've been using it in my daily work ever since. Last weekend I added two new agents focused purely on performance: a Unity Performance Reviewer and a Unity Profiler Analyst. And already yesterday, I got the first big win. On the very first try, the Profiler Analyst identified a real memory regression in our live game just by reading a Memory Profiler snapshot. The kind of issue that can easily slip past a normal code review but stands out immediately once something actually inspects what is resident in memory. I'll keep the specifics internal, but the takeaway was: a meaningful production-shape issue affecting scalability, found on input capture #1, on a codebase the agent had zero prior context on. What each of the two agents does: • Performance Reviewer reviews diffs through a pure runtime-cost lens: allocations in hot paths, GC pressure, draw-call growth, Canvas rebuilds, shader variant explosion, mobile frame budgets. It complements the generic code reviewer: that one cares about correctness, this one only asks "what will this cost at 60 FPS on a cheap device?" • Profiler Analyst doesn't touch code at all. It ingests Profiler / Profile Analyzer / Memory Profiler / Frame Debugger captures and returns a ranked, quantified optimization plan. That said, a grain of salt: one good catch isn't a pattern yet. Profile first, then optimize (if needed), even when the agent sounds very confident 🙃 Drop a 🔥 if you want more examples of how I use these at work. https://github.com/AlexMerzlikin/unity-agent-team #ai #agent #performance #unity

I am using Superpowers and I was surprised by its visual companion skill. Claude used it automatically when I asked it to introduce better character models into my prototype. Instead of only describing the plan in text, it suggested starting a local server and preparing a page with visual examples. Of course, it uses more tokens, but I feel like it made iteration on the new design much faster than before.

Anyway, the reason I got back to Antigravity was that I found an interesting repo I wanted to test: Agency Agents I checked the Antigravity docs and didn't find anything about agent setup. I just dropped them into a project like with Claude Code and asked the LLM to check it and fix the setup. It reassured me that everything was correct and that it supports agent teams. However, with all models, Antigravity fails to follow through the whole workflow. I have to pinpoint missed steps for every single task. Heavily inspired by the original repo, I prepared a team specialized in game development with Unity: https://github.com/AlexMerzlikin/unity-agent-team I also added custom tools to the Coplay Unity MCP to take screenshots and interact with the in-game logic. This way, the LLM can write manual test cases, run the game, take screenshots, and query the scene hierarchy to verify these "manual" tests. First impressions of these agents: • As said above Antigravity doesn't support it at all. • Codex and Claude Code integration is great. You can check what each agent does, and they run in parallel, making the whole flow a lot faster. • This team produces a lot of documentation. Compared to the default old Sonnet, which created multiple docs per prompt, this team does it better though. Everything is structured according to Nexus Sprint rules, and I feel like it helps to understand how your requirements transform throughout the dev pipeline. Also, each agent is super focused. Subjectively, there is less fluff in each document, and I get the exact info I need from a particular output. It's great for when I work on a big project, but it's too much overhead for prototypes. I prefer to skip it completely when just testing ideas. • Token drain. Each agent has its own context, so token spend multiplies. Keep that in mind when testing these agents, or tailor it to reduce the number of agents. If anyone tries out the Unity agent team repo, drop a comment. I am curious to see how it handles your specific setups and workflows. #ai #agent

Around 4–5 weeks ago I used my Google account with an AI Pro subscription to set up OpenClaw. Turned out it was against ToS and I was banned overnight from using Antigravity and Gemini CLI. I believe the correct option was to use the API key created in the Google Cloud panel. However, despite having monthly credits on my cloud account included in the Pro subscription, I couldn't use it unless I paid $30 directly on top, which is weird to me. I wrote an email to them, got no reply obviously, and forgot about Gemini completely until recently. I decided to check it again and I was unbanned. If you were banned too, is it working for you now? I don’t know if my email helped or they unbanned everyone.

Played around with Claude Code and remote controls while working on a prototype. Good alternative to doomscrolling. As you might know, the dopamine hit usually comes from the anticipation of finding a good video while scrolling, not the video itself. Now you can just "scroll" prompts to an agent instead. It’s useful for commuting when you don't have a laptop but still want to be kinda productive. The setup: - a web game (Three.js or the prototyping phase) - claude code - remote control (CC feature) - github action (publishes to itch.io draft on tagged master commit) The loop: I write a prompt, Claude implements it and tests via the Chrome Dev Tools MCP. If it reports success, I ask it to be pushed to master with a tag. In about a minute, the build is playable in a mobile browser on itch.io. I can play and iterate on it further from there. Can the anticipation of an agent’s output actually compete with a TikTok algorithm for you, or is this just a honeymoon phase?

Running Unity Tests at Warp Speed with .NET A while back, I posted about running Unity tests faster using .NET. So I sat down and documented the entire setup. https://gamedev.center/run-unity-tests-faster-dotnet/ #testing