Pinned Tweet
Here's the type of action you can expect from every single match in Impact Soccer. All clips taken from the same match. ⚔️
Game is live now! 👇
#ROBLOX
Game Dev Tip #7: The Compliance Agent for Roblox Devs
This might be the most underrated agent in my AI Agent Game Studio. If you don’t have one yet, you need to set one up ASAP as Roblox is going through a massive ban wave.
Give this prompt to your Orchestrator Agent:
Game Dev Tip #6:
Leveraging your AI Agent Game Studio on PC to Optimize for Performance on Console
When running your AI Agent Game Studio on your computer, your agents will not inherently have access to your Console. Your Game Editor may or may not have an emulator built in that can help with UI (Roblox studio does), but it won’t be able to accurately simulate performance.
When building on my PC, I have the QA agent run performance reports and that works great, but since it doesn’t have access to my Xbox, what can be done?
The good news is that there are multiple ways to set this up, but I went with a $0 cost option that requires the human in the loop. Essentially every time I play my live game on my account, regardless of platform, it turns it into a performance review.
Here is how I have it set up:
1. Build a LocalScript that ships with your game and records performance data, but only for your own account*. Here are the metrics my LocalScript reports:
- For every frame, record the frame time
- Every 30 seconds, determine the fps, p95 and worst frame time, physics step time and memory.
- Platform type (e.g. PC, Xbox, phone, etc.) and graphics setting.
*Note: it is important to only set this up for your account so that you are not running into data privacy issues.
2. Every 30 seconds, send a data summary to the server. The server will treat the data as untrusted: it will clamp the numbers, accept only allow-listed values and rate limit. The server will then add its own context to the data received - where the player was, player count, and its own tick rate.
3. The server saves the data to a DataStore under a single key for my account, keeping only recent data.
4. My agents read the data back through a Studio Playtest and build a table by place, setting, and device.
One of the first results I got back from my Xbox: the lobby holds 60 fps, but matches dip to 52–55 during collisions, and shadows made no difference. Physics time tracked fps almost one-to-one, so my agents are now fixing collision physics, not graphics.
Hope this is helpful 🤝
1
2
279
“Set up a compliance reviewer for my Roblox game(s) and run a first audit.
1. CREATE THE AGENT
Create a subagent called "release-compliance". Its rules:
- It is a reviewer. It never edits game code; it hands fixes to the agent or person who owns that code.
- It never submits forms, appeals or questionnaires. It drafts them and I submit.
- For every policy claim it opens the current official Roblox page (Community Standards, Terms of Use, and the Creator docs on content maturity, paid random items, text filtering, PolicyService and data stores), quotes the sentence it relies on and gives the link. Forum posts are labelled as opinion.
- It answers GO, GO WITH CONDITIONS or NO-GO, with file and line for each finding.
- Anything it could not verify is labelled NOT RUN. It never guesses.
- It writes each review to a dated report file and gives me a short summary.
2. ROUTE WORK TO IT
From now on, send it for review before I publish:
- every publish (admin and dev tools locked to the owner on the server, debug switches off, features under review still off)
- anything involving Robux: products, passes, wording, random rewards, trading, anything staked on chance
- anything that stores or sends player data
- anything players can type, name, build or upload that other players can see
- any social, chat, voice or private-space feature
- any change that could alter a maturity questionnaire answer
3. FIRST AUDIT, FOR EACH GAME IN THIS WORKSPACE
Read the code and report on:
- Child safety: is every piece of player-written text filtered before another player sees it, including text stored and shown later? Any links or prompts that send players off-platform? Any request for personal information? Any dating or romantic role-play? Are age and region restrictions read from PolicyService and respected?
- Maturity questionnaire: do my current answers match what is really in the game?
- Monetization: are odds disclosed for paid random items and are regional restrictions respected? Can a buyer lose or be misled about something they paid for? Anything that works like gambling?
- Player data: what is stored, for how long, and can I honour a deletion request?
- Admin and dev tools: is every check done on the server?
- Third-party code: free models, plugins, obfuscated scripts, require() by asset ID, HTTP calls to domains I do not control.
- Assets and metadata: music, images, brands, and a title, thumbnail or description that could mislead.
For each finding give: severity (blocker, should fix, note), the evidence, the policy quote and link, and the suggested fix. Do not change anything. End with what you could not check and what only I can decide.”
1
76
Times my compliance agent kept Impact Soccer out of trouble:
1. Paid gear with no warning (Nov 9)
In Impact Soccer when you hit another player in the game they lose one random piece of armor that can be equipped by the other team until the next goal is scored. Roblox prohibits removing functionality for buyers without proper notice. Fix: an info button on every Robux item and a one-time confirmation before the first purchase. It also told me an info button alone was not enough, because a buyer can pay without opening it.
2. No way to delete a player's data (Nov 8)
While reviewing new performance telemetry, it found the game could not honour a Right to Erasure request for player profiles. Fix: reworked storage keys and deletion templates.
3. Telemetry on a game minors can play (Nov 8)
It required the telemetry to be anonymous for players, numbers only, with 30-day retention. Its note: a 16+ audience still includes 16 and 17 year olds, so collecting less data matters more, not less.
4. Gambling and paid random items (Nov 4-5)
Before gear with gameplay effects and a lobby PvP mode shipped, it checked both against Roblox's definitions and held the rule that no paid item changes a random outcome.
5. Test hardware (Nov 8)
When I asked about automating console tests, it said controller emulator devices were a no.
6. A new admin command (Nov 10)
It confirmed that only the owner can reach it, had two risky switches blocked from being turned on live, and caught two log lines that printed a username.
70
Will be testing this out with my Mini AI Game Studio. Let’s see how it goes!
Introducing Claude Fable 5.2 and Claude Mythos 5.2, designed to tackle the most daunting technical challenges.
They’re capable collaborators on ambitious projects in many fields, and show improvements in areas like coding, reasoning, and vision.
1
60
Game Dev Tip #6:
Leveraging your AI Agent Game Studio on PC to Optimize for Performance on Console
When running your AI Agent Game Studio on your computer, your agents will not inherently have access to your Console. Your Game Editor may or may not have an emulator built in that can help with UI (Roblox studio does), but it won’t be able to accurately simulate performance.
When building on my PC, I have the QA agent run performance reports and that works great, but since it doesn’t have access to my Xbox, what can be done?
The good news is that there are multiple ways to set this up, but I went with a $0 cost option that requires the human in the loop. Essentially every time I play my live game on my account, regardless of platform, it turns it into a performance review.
Here is how I have it set up:
1. Build a LocalScript that ships with your game and records performance data, but only for your own account*. Here are the metrics my LocalScript reports:
- For every frame, record the frame time
- Every 30 seconds, determine the fps, p95 and worst frame time, physics step time and memory.
- Platform type (e.g. PC, Xbox, phone, etc.) and graphics setting.
*Note: it is important to only set this up for your account so that you are not running into data privacy issues.
2. Every 30 seconds, send a data summary to the server. The server will treat the data as untrusted: it will clamp the numbers, accept only allow-listed values and rate limit. The server will then add its own context to the data received - where the player was, player count, and its own tick rate.
3. The server saves the data to a DataStore under a single key for my account, keeping only recent data.
4. My agents read the data back through a Studio Playtest and build a table by place, setting, and device.
One of the first results I got back from my Xbox: the lobby holds 60 fps, but matches dip to 52–55 during collisions, and shadows made no difference. Physics time tracked fps almost one-to-one, so my agents are now fixing collision physics, not graphics.
Hope this is helpful 🤝
Game Dev Tip #5:
One of the critical components of the AI Agent-run Game Studio, is the QA agent. This computer use agent runs visual and log based tests in your game editor to confirm updates were completed correctly.
This is great for when you are adding new features or working out bugs, but during those testing sessions, the QA agent’s scope is limited to the particular task at hand.
In theory (assuming Separation of Concerns best coding practices are always followed by the coding agents), the QA agent tests should be enough, but in reality, updates sometimes cause bugs in other areas that the QA agent didn’t even think to look for.
To combat this, it makes sense to implement a “Smoke Test” as one last final QA Agent check prior to handing over to the human in the loop for manual testing.
The Smoke Test does the following:
- Ensures the code in the Game Editor matches your source
- Confirms every script loads with 0 errors, and the startup boots all systems into the expected state.
- Review that all dev and test features are turned off that shouldn’t be in production
- Conduct at least a 5 minute Playtest going through the core game loop, modes, menus, features, abilities, etc. to ensure everything is working properly.
Hope this is helpful 🤝
Fellow #Roblox devs:
If you are looking for play testers, let’s help each other out! I’ll test your game in exchange for testing my game. Feedback based only on the game itself. No judgement on tools used to help develop the game.
Please let me know in the comments if you are interested and we can start a DM 🤝
3
189
Dave - Roblox: Impact Soccer retweeted
A friend from our stream community shared something with me that I wanted to pass along.
On October 25th, a fire devastated Lulu M. Ross Elementary School in Milford, Delaware.
Thankfully, nobody was there when it happened.
But classrooms, supplies, and years of materials teachers and students depended on were lost or damaged.
Now the entire school district is coming together, moving grades, relocating classrooms, and doing everything possible to keep these kids learning.
And they could use our help.
I'm sharing the official recovery information below. Even if you can't donate, a repost could help this reach someone who can.
Let's show this community what we can do together.
It’s funny how bad AI is at predicting how long it will take to build something but everyone just gives it the pass.
I needed to optimize my game for console before flipping the switch on and it said it would take 3-5 days of AI agent time. It was done in a few hours 😂
1
57
Game Dev Tip #5:
One of the critical components of the AI Agent-run Game Studio, is the QA agent. This computer use agent runs visual and log based tests in your game editor to confirm updates were completed correctly.
This is great for when you are adding new features or working out bugs, but during those testing sessions, the QA agent’s scope is limited to the particular task at hand.
In theory (assuming Separation of Concerns best coding practices are always followed by the coding agents), the QA agent tests should be enough, but in reality, updates sometimes cause bugs in other areas that the QA agent didn’t even think to look for.
To combat this, it makes sense to implement a “Smoke Test” as one last final QA Agent check prior to handing over to the human in the loop for manual testing.
The Smoke Test does the following:
- Ensures the code in the Game Editor matches your source
- Confirms every script loads with 0 errors, and the startup boots all systems into the expected state.
- Review that all dev and test features are turned off that shouldn’t be in production
- Conduct at least a 5 minute Playtest going through the core game loop, modes, menus, features, abilities, etc. to ensure everything is working properly.
Hope this is helpful 🤝
Game Dev Tip #4:
Here are some recent improvements that were made to my AI Agent Game Studio:
- Phone Notifications: I noticed that I was losing productivity by being unaware that Claude was either waiting on a response from myself or when all the agent work was completed. Ask your highest level agent to send a notification to your phone (via /remote-control discussed in tip #3) when input is needed or work is completed.
- Night Shift Script: At least once everyday, I would come back to my computer or check the status of work on my phone, only to find out that my computer screen was locked and my QA agent (computer use agent) was unable to properly complete its test. I researched the different options built into windows and wasn’t finding the right option that would allow for my computer to remain on in cases when agents are working. So I had my Orchestrator agent run a simple script that flips on anytime agents are working that will keep my computer awake. Script get’s flipped off when work is done.
- Agent Approval to Save and Publish: I can recall multiple instances where Roblox Studio crashed over night and the updates were not saved. If I know I am going to be asleep when the updates finish, I will ask the Orchestrator agent to save to Roblox when the work is done so I can check in the morning. And in regards to publishing ability, obviously this is something you want to be careful about once you have a player base, but when I am not at my computer and want to test the updates on my phone, I can have the Orchestrator agent publish and I can complete my testing.
Hope this is helpful 🤝
New kick mechanic that controls your pitch based on how you position the camera.
Line drives = position camera to field level
Need more of an Arc = pitch the camera higher up
Lob/Lofty kick = pitch the camera to face the ground (best for teeing up headers to your team mates)
Let me know what you think!
2
277
I’m the idiot with the jester hat and puffy vest 😂. Watch me get cracked from behind with a ladder while the opposing team nails a header from 3/4 court… I must say this game gets my blood pumpin
Game Dev Tip #4:
Here are some recent improvements that were made to my AI Agent Game Studio:
- Phone Notifications: I noticed that I was losing productivity by being unaware that Claude was either waiting on a response from myself or when all the agent work was completed. Ask your highest level agent to send a notification to your phone (via /remote-control discussed in tip #3) when input is needed or work is completed.
- Night Shift Script: At least once everyday, I would come back to my computer or check the status of work on my phone, only to find out that my computer screen was locked and my QA agent (computer use agent) was unable to properly complete its test. I researched the different options built into windows and wasn’t finding the right option that would allow for my computer to remain on in cases when agents are working. So I had my Orchestrator agent run a simple script that flips on anytime agents are working that will keep my computer awake. Script get’s flipped off when work is done.
- Agent Approval to Save and Publish: I can recall multiple instances where Roblox Studio crashed over night and the updates were not saved. If I know I am going to be asleep when the updates finish, I will ask the Orchestrator agent to save to Roblox when the work is done so I can check in the morning. And in regards to publishing ability, obviously this is something you want to be careful about once you have a player base, but when I am not at my computer and want to test the updates on my phone, I can have the Orchestrator agent publish and I can complete my testing.
Hope this is helpful 🤝
Game Dev Tip #3:
Our approach to running a mini video game studio of AI agents should mimic that of an actual game studio. Human employees don't need to be told what to do every morning when they show up to work. They know their priorities and keep working off of them.
Here's how I replicate that with my own studio. Now I just type “claude start” and the team gets to work:
1. Turn on Remote Control for every session. In Claude Code, go to /config > "Enable Remote Control for all sessions." This is a key step: it's frustrating to realize you're dead in the water when you're out and about.
2. Add a startup hook that opens your tools. Have your Orchestrator agent (the highest-level agent) set up a SessionStart hook. It's a script that runs automatically every time a new Claude session starts, with no approval needed. Here is how I have mine setup:
A. launches Rojo in its own PowerShell window
B. opens Roblox Studio straight into the Impact Soccer experience
Note: It skips anything that's already running, so nothing gets restarted mid-work.
MCP Server Note: the Roblox Studio MCP turns on with a setting in Studio, then starts by itself whenever Studio opens, so I have no step for it. Other game editors may need a "launch and connect MCP server" step.
3. Put the morning routine in your CLAUDE.md file. This is what the Orchestrator does once everything is open:
A. Checks that Studio and Rojo are healthy, and has the QA agent click "Connect" on the Rojo plugin.
B. Catches up: it reads the last session's handoff notes, git and the plan.md file.
C. Resumes anything that was paused, then gives the top item in each lane (QA, building, art, audio/video, marketing) to the right agent.
D. Reports what's running, plus a list of things only I can test or decide.
The plan.md file is where all my brain-dumped ideas go. The Orchestrator prioritizes them automatically unless I give one a priority when I dump it. At the end of every session it writes a handoff and updates the queue, so the next session picks up exactly where we left off. Every once in a while it's still good to review plan.md yourself, to make sure everything is properly prioritized (and to remember all the ideas you forgot about lol).
Due to the limited RAM on my computer, I don't launch Blender or OBS at startup. Instead, I've authorized the Orchestrator to open and close them when needed. If you have a more powerful setup, you could add them to your startup hook.
Hope this is helpful 🤝
Roblox Devs:
Would you be interested in joining a group chat on X that focuses on helping play-test each other’s games? It seems there are more game devs now than play-testers so let’s help each other out.
Feedback will strictly be limited to the game itself. Criticisms over the tools used (or not used) to help develop the game will not be tolerated.
If this sounds like something you are interested in, please reply below. 🤝
New marketplace turned out great. Big map changes launching soon! The lobby will not be so safe anymore.. ⚔️
2
332
Ready for battle? ⚔️
roblox.com/share?code=2fca17…
119







