About a third of the English first requests Roblox developers typed into Picoo asked for an entire game, and only about 3 in 100 asked to fix something that was broken. We studied the opening requests of 744 conversations from 507 developers between 1 May and 18 October 2026; the rest mostly asked for one part of a game, a 3D build or map most often, then game logic, then UI.
Picoo is the Roblox Studio AI plugin we make, so this is first-party data: what people asked an AI to build, in their own words, before it built anything. Roblox says 44% of its top 1,000 creators now use Roblox Assistant or third-party AI tools (Roblox, April 2026), and its new Build tool turns a text prompt into a basic game (Roblox, July 2026). The question this study answers is what developers ask those tools for.
Key findings
- A third ask for a whole game. 32.2% of English first requests (127 of the 395 we coded) asked for an entire game, from a few words to design documents thousands of words long.
- 3D builds and maps lead the requests for one part. Of first requests that asked for one part of a game, 36.7% were 3D builds, maps or visual changes, 23.1% game logic, 15.7% UI, and 24.5% asked for several of these at once.
- Fixes come later. 3.3% of English first requests asked to fix something. Later on, 17.5% of conversations and 11.3% of follow-up requests used a fix word such as fix, bug, not working or crash; bug reports written without those words are not counted, so the real shares are higher.
- Length varies enormously. The median first request was 43 words, but 23.5% were under 10 words and 18.9% were over 300. 23.3% read like a written spec, with headings, bullet lists or a line such as "You are an expert Roblox developer".
- Simulators and tycoons are the genres named most. Among requests for a whole game, 72.1% named a genre: simulator (13.5%), tycoon (12.5%), horror and fighting or PvP (10.1% each) and brainrot or "steal a" games (9.1%) led.
- UI comes up in 4 of 10 conversations. Counting every request in a conversation, 40.8% mentioned UI or GUI, 32.8% maps, 29.4% animation, 24.1% shops and 22.7% coins or currency. DataStore saving came up in 10.3%.
- About 1 in 7 first requests is not in English. 85.3% were in English; Spanish (4.6%), Russian or another Cyrillic-script language (3%), French (1.6%), German (1.3%) and Portuguese (1.2%) came next.
The sample
The data covers every Picoo conversation created between 1 May and 18 October 2026 (UTC) by an account that is not our own. After removing our own and test accounts, deleted accounts, example prompts that people sent with one click, button presses and automatic "continue" messages, 766 conversations from 520 developers contained at least one request a person typed. 744 of them opened with a typed request, and 635 of those were in English. The methodology section at the end lists every rule and count.
Most developers appear only once or twice: 310 of the 428 English-writing developers contributed a single first request, no one contributed more than 10, and the busiest 5% of developers account for 17.5% of first requests.
Whole games, 3D, scripts or UI: how first requests split
We read a random sample of 395 of the 635 English first requests and put each into one group (the methodology section at the end has the definitions):
- An entire game: 32.2% (127)
- A 3D build, map or visual change only (a map, a building, models, terrain, lighting): 21.3% (84)
- Several parts at once (for example a shop with its UI and the code that saves purchases): 14.2% (56)
- Game logic or scripts only (combat, abilities, tools, NPC behavior, saving): 13.4% (53)
- UI only (menus, HUDs, health bars, shop screens): 9.1% (36)
- Not a request to build anything (questions, connection or account problems, "read my game"): 9.9% (39)
With a sample this size, the margin of error at 95% confidence is at most about 5 percentage points for any of these shares (about 3 if these 635 English first requests are treated as the whole population). That margin covers sampling error only, not coding error. Counting one request per developer gives almost the same split (35% whole games among 303 developers), so a few heavy users do not drive it.
The groups differ a lot in length. The median request for a whole game was 137 words, a request for several parts 84 words, a 3D request 40.5 words, a scripting request 40 words and a UI request 15.5 words.
The genres people name when they ask for a whole game
208 English first requests from 168 developers asked for a whole game by our keyword rules, and 150 of them (72.1%) named at least one genre. A request can name several:
- Simulator: 28 requests (13.5%), from 26 developers
- Tycoon: 26 requests (12.5%), from 21 developers
- Horror: 21 requests (10.1%), from 18 developers
- Fighting, PvP or battlegrounds: 21 requests (10.1%), from 21 developers
- Brainrot or "steal a" games: 19 requests (9.1%), from 15 developers
- Survival: 18 requests (8.7%), from 14 developers
- Obby or parkour: 15 requests (7.2%), from 12 developers
- Round-based or minigames: 13 requests (6.2%), from 9 developers
- RPG or adventure: 12 requests (5.8%), from 12 developers
- Clicker or idle: 12 requests (5.8%), from 11 developers
- Escape games: 10 requests (4.8%), from 8 developers
- Sports: 9 requests (4.3%), from 8 developers
The other 58 named none of these genres. Counts this small move easily, so read the order of the lower entries as rough.
Our July count of every message found genre words rare. That count divided by every message sent through July, including one-word replies such as "continue" and "go" and our own test messages. Counted per first request, tycoons and simulators are what people name most when they ask for a whole game. If you are starting one, the simulator and tycoon guides show how those loops are put together.
The game parts that come up across a conversation
Later requests add parts the first one did not mention, so this count looks at every typed request in each of 652 English conversations (from 435 developers) and counts a part once per conversation:
- UI or GUI: 40.8% (266 conversations)
- Maps: 32.8% (214 conversations)
- Animation: 29.4% (192 conversations)
- Shops: 24.1% (157 conversations)
- Coins or currency: 22.7% (148 conversations)
- Sound or music: 21% (137 conversations)
- Lighting or atmosphere: 18.9% (123 conversations)
- Particles or visual effects: 18.6% (121 conversations)
- Abilities or powers: 17.6% (115 conversations)
- Inventory: 16.3% (106 conversations)
- Combat or damage: 15.8% (103 conversations)
- Cutscenes or camera: 15.3% (100 conversations)
- Enemies, bosses or zombies: 15% (98 conversations)
- Menus or loading screens: 14.7% (96 conversations)
- Rebirths, levels or XP: 13.3% (87 conversations)
- NPCs: 12.9% (84 conversations)
- RemoteEvents: 12.9% (84 conversations)
- Mobile or touch controls: 12.3% (80 conversations)
- Teleporting: 11.2% (73 conversations)
- DataStore saving: 10.3% (67 conversations)
- Leaderboards: 10.1% (66 conversations)
- Pets: 7.7% (50 conversations)
UI leads by a wide margin, and presentation work (animation, sound, lighting and effects) shows up more often than core systems such as inventory, combat or saving. Shops (24.1%) and coins or currency (22.7%) each appear in about 1 in 4 conversations. See the menu UI and shop guides for how those are usually built.
How long first requests are
The 635 English first requests spread from a single word to several thousand:
- 1 to 5 words: 11.2%
- 6 to 15 words: 20.8%
- 16 to 40 words: 16.7%
- 41 to 100 words: 15.9%
- 101 to 300 words: 16.5%
- More than 300 words: 18.9%
The median was 43 words; a quarter were 10 words or fewer and a quarter were 211 words or more. 148 requests (23.3%) were longer than 40 words and formatted like a spec: markdown headings, bold labels, numbered or bulleted requirement lists, or a line such as "You are an expert Roblox developer".
Fix or build: what people ask for first and later
Of the 395 coded first requests, 78.5% asked for something new, 9.1% asked to change or improve something that already existed, 3.3% asked to fix something broken, and the other 9.1% were questions, account or connection problems, requests to look at the game, or no request at all.
Fixing grows as a conversation goes on. Of 652 English conversations, 114 (17.5%) included at least one request with a fix word such as fix, bug, not working or crash, and 11.3% of the 2,907 follow-up requests used one. These are keyword counts, not coded requests, so they are lower bounds: a bug report written without one of those words is not counted, and the real share of fix requests is higher.
Conversations are short. Across all 766 conversations, the median had 2 typed requests and 32.9% had only one; a minority of long sessions pulls the mean up to 5.3.
What this means if you build with AI
- A one-line request for a whole game leaves every decision to the AI. If you care about how the shop, saving or controls work, say so up front or ask for them one at a time.
- Name the parts players touch. UI comes up in 4 of 10 conversations, and mobile controls and saving in about 1 in 10. If your game needs them, ask for them in the first request rather than as an afterthought.
- Plan for a second pass. At least 1 in 6 conversations included a request with a fix word such as bug or not working, so test early, in Studio, before adding more on top. Our guide to using AI in Roblox Studio covers how to check what an AI changed.
How Picoo fits in
These requests were typed into Picoo, an AI copilot for Roblox Studio: you chat in your browser, and its Studio plugin builds the result into your place. Its how-to library shows how common systems such as obbies and DataStore saving are built. See how it works at picoo.io.
Methodology
Data and window
We read Picoo's own records (read-only) on 19 October 2026: the conversations table and every message a user sent. The window is conversations created from 1 May 2026 to 18 October 2026, UTC; the first conversation in the table is from 1 May. 1,237 conversations were created in that window.
Who was left out
- Our own and test accounts: 319 conversations from accounts that Picoo's code treats as its own (company email domains and the founder's accounts).
- Deleted accounts: 112 conversations from 8 accounts that no longer exist, because we could not check whether they were ours. One of them had sent our own test prompts dozens of times.
- Messages nobody typed: 691 automatic "continue" messages that the app sends between parts of a long build, and 64 example prompts sent with one click: the example buttons on the chat page (current and past versions), the copy-and-paste examples on the dashboard, and card buttons such as "Try again". A message counted as an example if it matched one word for word, ignoring case and spacing, or started with the same 40 characters.
- Controls: 1,077 short messages made only of words like "go", "continue", "yes", "ok", "retry" or "hi", or a lone command such as "/go". Some are button presses, some are typed; none says what to build.
That leaves 4,022 typed requests in 766 conversations from 520 developers. The first request of a conversation is its first typed request. The 57 conversations that opened with an example prompt are left out of the first-request numbers; their later typed requests still count in the conversation numbers.
Language
A request counts as English when English function words (the, and, to, with, my...) outnumber those of any other language we checked, or when it is plain ASCII with no language markers at all ("tycoon", "Simulator"). Requests written mostly in another script (Cyrillic, Arabic, Thai, Chinese, Korean and others) are labeled by script. 635 of the 744 first requests (85.3%) were English. Topic and genre counts use English keywords, so they are computed on English requests only.
Coding the request types
For the split into whole games, 3D, scripts, UI and mixed requests, and for the build, change or fix split, we drew four random samples of 100 English first requests, each from the requests not drawn before, and coded each request by reading its first 420 characters, the whole request for most of them. Five drawn requests were later reclassified as example prompts, a control message or non-English, which leaves 395. The coding was AI-assisted: an AI model read each request and applied these definitions.
- Entire game: asks for a whole game or game mode, by genre ("a tycoon"), by concept ("a game where...") or as a design document.
- 3D only: models, maps, buildings, terrain, environments, lighting or visual effects, with no new game logic or UI.
- Scripts only: game logic, systems, tools, abilities, NPC behavior or saving, with no new UI or 3D build.
- UI only: screens, menus, HUDs, bars, buttons or shop interfaces.
- Mixed: clearly asks for two or three of the above.
- Not a build request: a question, a connection or account problem, a request to look at the game, or a message with no ask.
- Fix: asks to repair something broken or not working. Change: asks to modify, improve, move or remove something that exists. Build: asks for something new.
As a check, we also wrote keyword rules for the same groups and ran them on all 635 English first requests. On the last of the four samples, coded after the rules were frozen, the rules agreed with the coding on 86 of 100 requests for whole game or not, and on 99 of 100 for fix or not, and they put 32.8% of all English first requests in the whole-game group, against 32.2% in the coded sample. They agreed on only 66 of 100 for the full six-way split, which is why the split above comes from the coded sample. Most disagreements were about whether a request that names several parts counts as mixed or as its main part, and about whole game versus one part.
Genres and game parts
Genres and parts are keyword matches, case-insensitive, with a request (or conversation) counted once per genre or part. Genre examples: simulator; tycoon; horror, scary game, backrooms; PvP, fighting game, battlegrounds; brainrot, "steal a ..."; survival; obby, parkour, platformer; minigames, round-based; RPG, open world, adventure game; clicker, idle game. Part examples: UI, GUI, ScreenGui; map or maps; animat...; shop or shops; coin, currency, cash; sound, music; lighting, atmosphere; particle, VFX, trail, explosion; inventory; RemoteEvent or RemoteFunction; NPC; DataStore or DataStoreService; mobile, touch controls, joystick; teleport; leaderboard. The whole-game group for the genre list comes from the keyword rules, since the coded sample is too small to split by genre.
Fix requests across a conversation
A typed request counts as a fix when it contains words such as fix, bug, not working, doesn't work, broken, glitch, crash or debug. A request of more than 60 words is judged on its first 200 characters, because long specs list "no bugs" or "handle errors" as requirements, and a fix word right after words such as no, without, never, avoid or prevent does not count. These are keyword matches, not coded requests: a bug report that uses none of these words is not counted, so the conversation and follow-up shares are lower bounds.
Length
Length is counted in words, meaning runs of characters between spaces, on the 635 English first requests.
Limitations
These are requests from developers who chose an AI plugin and signed up for Picoo, not a survey of all Roblox developers. Picoo's own features, such as its recipes for common systems, may shape what people ask for. Keyword counts miss synonyms and typos, and the AI-assisted coding was not double-checked by a second coder. All numbers are aggregates; no request is quoted. Questions about the data: help@picoo.io.
FAQ
What do Roblox developers ask AI to build most?
In this study, about a third of English first requests asked for an entire game, most often a simulator, tycoon, horror or PvP game. Among requests for one part, 3D builds and maps were most common, then game logic, then UI. Across whole conversations, UI came up most, in 40.8% of them.
How long is a typical first request to an AI for Roblox?
The median was 43 words, but requests split into short asks and long specs: 23.5% were under 10 words and 35.4% were over 100 words. Requests for a whole game were the longest, with a median of 137 words.
How often do developers ask AI to fix code instead of building something?
Rarely at first: 3.3% of English first requests asked for a fix. Later it is common: 17.5% of conversations and 11.3% of follow-up requests used a fix word such as fix, bug, not working or crash. Bug reports written without those words are not counted, so the real shares are higher.
Can I use these numbers?
Yes. Please cite Picoo and link to this page. Every number is an aggregate over the sample described in the methodology, dated October 2026.