AI & IT toolsAI Workflow

Splitting solo game dev work with AI: what to hand off, what to check, terms, and Steam and itch.io AI disclosure

When you make a game alone, how much of the code, 3D assets, art, sound, and text can you hand to AI, and where must a person check the result? This guide covers the commercial-use terms of each kind of tool, the AI disclosure rules on Steam and itch.io, and what AI actually did in a Godot game test.

AZTechnology18 min read

한국어English

Based on: Claude Code, Blender 5.2, Godot 4.7.2, each tool's terms of use, and the official Steamworks and itch.io guidance. Checked on November 4, 2026. This is not legal advice. Terms change, so check them again before you rely on them. Figure 1 is a real game screen. Figures 2 and 3 are official web pages. The red numbers and boxes were added for explanation.

When you make a game alone, the work splits into five areas: code, 3D assets, art, sound, and text. Today there are AI tools for all five. So it is easy to think "I can just have AI do everything." In practice, how much you can hand off is different in each area, and so is the place where a person has to check. On top of that, if you want to sell the game or put it on a store, you must follow each tool's terms and the store's disclosure rules.

This is part 4 of the series "Solo game development with Blender, AI, and a game engine." It goes through what to give AI and what to check yourself in each area, the commercial-use terms of the tools, and the AI disclosure rules of stores. It also lists what AI actually did in a test game I made with Godot. The whole series starts at the solo game dev map.

1. The short answer: each area at a glance

AreaGood to hand to AIA person mustTerms and rules to read first
CodeGame rule code, export scripts, automatic checks, bug fixesDecide what to make, play it and say how it feelsThe output clause of the coding tool's terms, the engine license notice
3D assetsBuilding shapes with Blender Python scripts, repetitive work, exportingCompare size, direction, color, and names with numbers and picturesOwnership by plan for 3D generation tools
Art (textures, 2D)Drafts, patterns, icon sketchesCheck that it does not look like someone else's work and that text is not brokenImage tool terms, store disclosure rules
SoundDraft sound effects, sounds made in codeListen and judge whether it is annoyingCommercial use by plan for sound tools
Text (UI text, descriptions, translation)UI text, help text, draft translationsCheck facts and tone, decide what goes inFont license, store disclosure rules

In one line: AI is the hands that make and fix; the person is the eyes that decide and check. In house building, AI is the fast bricklayer. Choosing the design, then living in the finished house and finding what to fix, is the owner's job.

2. A real record: who did what in a Godot game test

On October 10, 2026, I made a small air defense game with Claude Code and Godot 4.7.2. Claude Code is an AI coding tool that runs in a terminal. It reads and writes the files in a folder and runs commands. Here is who did what in each area, limited to what the records show.

Air defense game screen in a low-poly desert at sunset. Box 1 is the top-left text: launchers left 4/4, a green bar, missiles left 18, and missile slots for launchers 1 to 4. Box 2 is white smoke and an orange explosion from a missile rising from the ground. Box 3 is a sand-colored launcher truck at the bottom left. Red diamond markers, a three-ring aim circle, and 80% text are in the sky Figure 1. The air defense game. (1) On-screen text (drawn by code that AI wrote) (2) Smoke and explosions (effects made in code) (3) Launcher model (made in Blender and moved over) (Godot run screenshot)

AreaWhat AI didWhat the person did
PlanDrafted the plan (PRD): three questions the test answers, pass criteria, what is out of scope, game rulesRead it and said "go with the proposal"
CodeGame code, a script that builds terrain, sky, and mountains in code, 14 automatic checks, one per rulePlayed it five times and said what to change. On the fifth: "difficulty is about right"
3D assetsA script that exports 8 Blender models to glTF, a size comparison between Godot and Blender (all differences 0)No record (received the comparison pictures and size table)
ArtNo image files used. Colors came from model materials; sky and ground from codeNothing
SoundSix sound effects synthesized in code when the game starts. No sound files and no paid toolsDid not comment on sound
TextAll on-screen text (launchers left, missiles left, and so on) in KoreanSaid "just remove the score," so the score display went away

The models were low-poly models made earlier for videos with Blender Python scripts, and those scripts were also written with Claude Code. Before that there were versions made with an AI 3D generation tool, but this test did not use them. The plan said not to use the folders holding those old versions.

The first playable version took about one hour. In the same session, a rigged character test and a penalty kick game brought the total to about three hours. Up to right after the rigging test, Claude Code used about 160,000 output tokens, or about 11 US dollars in cost terms. The longest part was not the code. It was the time I spent playing and saying what to change. The full process is in Making a Godot game scene with Claude Code.

What the table shows is that "checking" looks different in each area. For code, even when every automatic check passed, only a person playing could tell whether it was fun. For assets, comparing numbers was enough. For sound, I never listened and judged on purpose, so this record cannot say whether the sound is good.

3. Code: the area you can hand off the most

Code is the best area to hand to AI. The reason is that a machine can check the result. If each rule has a check such as "firing creates a missile" or "one drone hitting a launcher destroys it," AI can run the checks after every change and fix any FAIL. It is like giving homework together with the test and the answer key.

Follow along:

  1. Before asking for code, have AI write a one-page plan. It should state what to make, when it is done, and what it will not do.
  2. Have AI make one automatic check for each rule.
  3. Play one round yourself and say what to change in numbers. "Spawn twice as fast" gets tuned faster than "make it harder."
  4. Have AI update the plan when a rule changes.

There are two places a person must not miss. First, fun and difficulty cannot be measured by automatic checks. Second, check that AI did not loosen the checks themselves. If the number of checks suddenly drops, ask why.

Read the terms once, too. Anthropic's Consumer Terms of Service (effective October 8, 2025) say that, as long as you follow the terms, the company assigns its rights in outputs to you. The Commercial Terms used for the API and team plans (effective June 17, 2025) also say the customer owns its outputs. The same consumer terms tell you not to rely on outputs without checking their accuracy yourself. Getting the rights and having a correct result are two separate things.

The engine has its own duty. Godot uses the MIT license. When you ship a game, you must put the Godot license text somewhere the player can see it (a credits screen, a licenses menu, a file shipped with the game, and so on). This is the same whether AI or a person wrote the code.

4. 3D assets: ways to make them and ways to check them

There are two main ways to make 3D assets with AI.

WayHowStrengthWhat to check
Let AI drive BlenderAI writes Blender Python scripts or sends commands to Blender through MCPThe result stays as a script you can run again. Sizes can be set as numbersSize, origin, and names follow the rules
3D generation toolsGive a picture or text and get a modelShapes appear fastFace count, split vertices, ownership by plan

MCP is a connection standard that lets AI use another program as a tool. Like a power socket that takes any appliance with the right plug, any program that supports MCP can be driven by AI. How to connect is in Connecting Blender and AI, and what to check before handing over Blender is in A safety check before letting AI drive Blender.

When you use a 3D generation tool, always check the plan. For example, Meshy's terms of use (updated September 19, 2026) say Meshy owns the output made on the free plan and lets free users use it under the CC BY 4.0 license as long as they credit Meshy. They say users on a paid plan own their output. So if you put a model made on the free plan into a game, you must credit the source. How to clean up a generated model for a game is in Cleaning up an AI-made 3D model in Blender.

Blender itself has no limits on commercial use. The license page on blender.org says what you make with Blender (images, videos, and .blend files) is yours.

The person checks with numbers. In my record, I had AI compare the size of all 8 models in Godot against Blender, and every difference was 0. Even so, there were five places where colors went missing or a floor plate came along with the model. These were checked with the comparison pictures and size table, and the export script handled all five. The rules a game asset should follow are in the next part, Game asset rules.

5. Art, sound, and text: where terms differ the most

Art, sound, and text are what players see and hear directly, so they fall right under the store disclosure rules (section 7). It is also common for commercial use to depend on the plan.

Art. My two test games had no image files at all. Model colors came from materials, and the sky and ground were made in code. If you make textures or 2D art with AI, find the commercial-use and ownership clauses in that tool's terms and write them down. The person's job is to check that the art does not look like another work or a trademark and that any text in the picture is not broken.

Sound. The six sound effects in the air defense game (launch, explosion, launcher hit, when firing is not possible, when you win, when you lose) are computed in code as sound waves when the game starts. With no sound files and no paid tools, there were no sound terms to read. If you use an AI sound tool, check the plan. For example, on the ElevenLabs pricing page, the Commercial License appears from the Starter plan, and it is not in the free plan's list.

Text and fonts. AI wrote the on-screen text in Korean inside the code, and I decided what to show. I removed the score from the air defense game and the probability numbers from the penalty kick game, because a screen full of numbers bothered me when I played. In the canyon flight game, exported for the web, Korean text only showed after the font was put inside the game. A game running in a browser cannot use the fonts installed on the PC. I used a reduced font made from the free font Noto Sans KR, keeping only 2,350 Korean characters (424 KB). Its license, the SIL OFL, allows bundling the font with software and selling it with the software, with three conditions:

  1. The font cannot be sold by itself.
  2. Every copy you distribute must include the copyright notice and the license text.
  3. A modified font cannot use the original font's reserved names.

A font with characters removed also counts as a modified version. A modified version cannot use the original font's reserved name as its font name (the name stored inside the font and shown in font menus). The file name has nothing to do with it. The reserved name for Noto Sans KR is "Source." Check the first line of the font's license to find its reserved name.

6. Rights: how far "the output is yours" goes

The Anthropic terms in section 3 say the company assigns its rights in outputs, with the words "(if any)." Meshy's ownership clause for paid plans also adds "to the extent possible under applicable law." That means the company passes on whatever rights it has. It does not promise that copyright exists in the output.

Whether copyright exists is up to each country's law. The U.S. Copyright Office published a report (Part 2) on the copyrightability of generative AI outputs on January 29, 2025. The report concludes that, with today's technology, prompts alone do not give users enough control over the result to be treated as its authors, and it states that the parts a person creatively selected, arranged, or changed are protected. Copyright basics for AI output are covered separately in Copyright basics for AI-made images and text. For a solo developer, this means: keep a record of what you selected and changed, rather than using AI output as is. If a sale or a contract depends on it, ask an expert in that country's law.

7. Store rules for AI-generated content

Read the rules of the store you plan to use when you start, not after the game is done. Answering "was this art made with AI?" later needs a record.

The 3) Generative Artificial Intelligence Content section on the Steamworks Content Survey page. Box 1 is the paragraph saying efficiency gains from tools are not the focus and that the section is about content shipped with the game and consumed by players (artwork, sound, narrative, localization, etc.). Box 2 is the Pre-Generated and Live-Generated items Figure 2. The generative AI part of Steam's Content Survey. (1) What the section covers (2) Pre-Generated and Live-Generated (Steamworks documentation screenshot)

Steam. Before you submit a game for review, you fill in the Content Survey. Its third part is about generative AI. As (1) in Figure 2 shows, this part is not about efficiency gains from development tools. It asks whether you used AI to make content that ships with the game and that players see and hear (artwork, sound, narrative, localization, and so on). The two kinds in (2) are:

KindMeaningWhat you also do
Pre-GeneratedContent made with AI tools during development and shipped in the gamePromise there is no illegal or infringing content and that the game matches its marketing. Review judges it by the same standard as non-AI content
Live-GeneratedContent made by AI while the game runsAll of the above, plus describe the guardrails that stop it from generating illegal content

The Accurately tag your use of generative AI section of the itch.io quality guidelines. Box 1 is the paragraph asking creators to tag projects that contain generative AI material using the AI Disclosure section of the edit page. Box 2 is the paragraph saying self-contained algorithms such as pathfinding, procedural level generation, and difficulty adjustment do not need disclosure. Box 3 is the box saying disclosure is strictly enforced for game asset, soundtrack, physical game, comic, and book pages, and that failing to tag may result in delisting Figure 3. The generative AI part of the itch.io quality guidelines. (1) Tag it in the AI Disclosure section (2) Algorithms that need no disclosure (3) Where it is strictly enforced (itch.io documentation screenshot)

itch.io. If your project contains material made with generative AI, mark it in the AI Disclosure section of the project's edit page ((1) in Figure 3). Projects that do so also appear on the AI Assisted browse page. As (2) says, self-contained algorithms that do not use large external datasets (enemy pathfinding, procedural level generation, automatic difficulty adjustment, and so on) are not treated as generative AI and need no disclosure. As (3) says, disclosure is strictly enforced for game asset, soundtrack, physical game, comic, and book pages, and failing to tag may get a page delisted.

Neither page has a sentence that requires disclosure just because AI helped write code. Still, some cases are unclear. For example, a model built by a Blender script that AI wrote could be judged either way as "content made with the help of AI tools." When in doubt, it is safer to describe plainly what you used and how. My two test games were never put on a store, so I have no record of actually filling in these surveys.

8. Follow along: one record file to start with

Starting point: You have just made the game's project folder.

Follow along:

  1. Make one record file in the project folder, such as credits.md.
  2. Write one line per tool: tool name, plan, date you checked the terms, commercial-use conditions, attribution duty.
  3. Write one line per thing that goes into the game (models, art, sound, fonts, text): file name, how it was made (by hand, which AI tool, where it came from), and what a person changed.
  4. When you give AI work, write in the plan: "When you add a new asset, add a line to credits.md."
  5. Before exporting, use this file to fill in the credits screen and the store survey. Add the Godot license text at this point too.

For example:

What went inHow it was madeConditions
Launcher modelBlender Python script (written by AI), sizes from the planNone (made by me)
Six sound effectsSynthesized in game code (written by AI)None
Game fontNoto Sans KR reduced to 2,350 Korean charactersSIL OFL: include the license text, do not use the reserved name (Source)
EngineGodot 4.7.2MIT: make the license text visible to players

With this one table, you know right away what to write in the store survey and whom to put in the credits. It works like the contents list you stick on each box when you move house.

9. Common mistakes and fixes

SymptomCauseFix
Every automatic check passes, but the game is not funCheck scripts cannot measure funPlay it yourself and say what to change in numbers
The number of checks quietly went downAI changed the checks to get rid of a FAILRecord the check count and ask why if it drops
You forgot to credit a model made on a free planYou did not check the conditions for that planWrite the ownership clause in credits.md and credit the source
Korean text shows as boxes in the web buildA game in a browser cannot use PC fontsPut a free font inside the game and include its license text
Right before release, you cannot tell which art was made with AINothing was written down at the timeFrom the start, write one line per asset on how it was made (section 8)
The exported game has no engine license textYou did not know the MIT notice dutyPut the Godot license text in the credits screen or a shipped file

10. Summary

AI is the hands that make and fix; the person is the eyes that decide and check. Check code with automatic checks, assets with numbers, and fun and sound by playing and listening. Read tool terms and store rules at the start, and write one line on how each thing in the game was made.

In part 5, Game asset rules, we check polygons, textures, origin, scale, names, LOD, and collision shapes for models sent from Blender to a game engine, using real screens.

References

KEEP EXPLORING

Read next

Four frames from the first version of the air defense game. Launchers in a low-poly desert at sunset, the score, base health 100, and a green bar at the top left, time left 0:30, 0:17, 0:07, and 0:00 at the top right, and in the last frame the Korean text for Defense succeeded! Score 1500, R: restart (Godot run screenshot)AI & IT tools
AI Workflow13 min read

Setting the scope of your first game: a size a solo developer can finish, and how to schedule it

To take a first solo game all the way to the end, you need to cut it down to the right size. This guide covers what to write on a one-page plan and how to split the schedule into steps, with real times, snags, and scope changes from making three small games in Godot.

A low-poly desert at sunset with four launchers. A missile rises from the left launcher, a three-ring aim circle with a yellow 90% sits among red diamond markers in the sky, and the top left shows remaining launchers 4/4 and remaining missiles 31 (Godot run screenshot)AI & IT tools
AI Workflow13 min read

A map of solo game development: what Blender, AI, and a game engine each do, and the steps to your first game

When you make a game alone, Blender, an AI assistant, and a game engine each take a different part of the work. This guide sorts out who does what, lays out seven steps to a first finished game, and shows notes from making small games in Godot. It is the starting page of a five-part series on solo development.

The Script screen of the Godot editor. Under func _steer at line 102 of missile.gd are var to_t := target.position - position, var dist := to_t.length(), var want := (lead - position).normalized(), and var ang := dir.angle_to(want), with a list of functions on the leftAI & IT tools
AI Workflow16 min read

What helps to know before solo game development: what you need without a degree and what AI fills in

When you make a game alone, which parts of programming, math (vectors and coordinates), art, sound, and game design do you really need, and which can AI and tools fill in? This guide answers with real game code and notes, including what AI cannot fill in, and points to the official learning material for each engine.

Two fighter jet models and one small 3D character printed in white, on a gray stone-pattern tableAI & IT tools
AI Workflow24 min read

3D Printing from One Picture with Meshy AI and Claude Code: Printing a Fighter Jet Model and a Character

The real workflow for turning one reference picture into a 3D model with Meshy AI, checking and repairing it with Claude Code and Blender, and printing it on a 3D printer.

Top of the Artificial Analysis text-to-image leaderboard AA-Image-T2I v2.1. Rank 1 GPT Image 2.5 Sunburst (max) 1222, rank 2 GPT Image 2.5 Flare (max) 1207, rank 3 GPT Image 2 (high) 1186, rank 4 Nano Banana 2.1 1176, rank 5 MAI-Image-2.6 1166, rank 6 Grok Imagine Image 2.0 1166, rank 7 Nano Banana 2 1140, price per 1,000 images on the rightAI & IT tools
AI Tools13 min read

Choosing an AI image model in autumn 2026: GPT Image 2.5 and Nano Banana 2.1 compared

Names, release dates, prices, and leaderboard positions of GPT Image 2.5 (Flare and Sunburst) and Nano Banana 2.1, the image models that arrived in autumn 2026, checked against official sources, plus simple rules for picking one by use.

Chart comparing graphics memory use as horizontal bars. Screen only 754 MiB, AI model loaded 3,946 MiB, AI plus EEVEE render 5,162 MiB, AI plus Cycles GPU render 5,721 MiB, AI on CPU plus Cycles 2,618 MiB, card total 8,188 MiBAI & IT tools
Computer Hardware17 min read

Running Blender and a local AI on one graphics card: measured VRAM sharing on 8 GB

On my laptop with an RTX 4070 Laptop GPU (8 GB), I kept a local AI model (Gemma 4 E4B on llama.cpp) running while rendering with Blender EEVEE and Cycles. I measured how full the graphics memory got, how much the render and the AI slowed each other down, and whether to put the AI on the graphics card or the CPU. The post ends with steps to plan memory on your own PC.