Magnus Tech Cut the Path from Idea to Game Test from Weeks to a Couple of Hours
Magnus Tech stopped writing multi-page specs to validate ideas. The studio connected Claude Code to Unity through an MCP server, and the agent builds a rough prototype from primitives on its own. The path from concept to first playtest shrank from weeks to a few hours. A producer with no coding skills built a playable ad for a Match-3 game in 3 calendar days, including 4–5 hours of work with the agent. The studio keeps the core gameplay, architecture, and balancing in human hands.
AI-processed from Habr AI; edited by Hamidun News
In game development, an idea is cheap, but testing it by hand is expensive. The classic cycle is familiar to everyone: a few days go to a design document, then programmers spend a week or two writing code, and only after the build is running does it turn out the game is boring. Magnus Tech decided to cut the most expensive link from this chain. Instead of text specs running dozens of pages, the studio builds a rough prototype from primitives in a couple of hours and tests it right away. According to Ivan Dvoryanov, the game division's producer, the time from idea to first test has dropped from weeks to a few hours.
How the agent gets inside Unity
A chatbot in a browser tab is no good for this kind of work. The repository holds thousands of scripts, prefabs, and configs, the context window fills up fast, and copying text between the IDE and a chat gets tedious. So the studio connected Claude directly to the engine via MCP, a protocol that gives a model access to the tools of an external program.
The working setup looks like this: Claude Code inside VSCode and the open-source unity-mcp server (the MCP for Unity package) from CoplayDev. The agent does not try to cram the whole project into context. Instead, it selectively searches for the scripts it needs, reads scene structure and the GameObject hierarchy, works through prefab dependencies, and calls editor commands.
Installation takes a few steps: the package is added through the Package Manager, then Window → MCP for Unity is opened, Claude Code is selected, and auto-configuration is run. According to the authors, Claude Code can set all of this up itself if asked. Similar MCP servers already exist for Unreal Engine, Godot, Blender, and Figma.
For complex tasks, the studio raises the effort parameter, which gives the model more time to reason, and pays for it in tokens.
On one side project in Godot, the producer never opened the editor to assemble scenes. He entered the engine only as QA: he ran the build and checked how what the agent had made from his requests actually played.
Plan Mode instead of "just make it feel good"
The studio's main technique is Plan Mode. The producer describes a feature in plain words and attaches references. The agent studies the code through MCP and produces a technical plan of roughly one A4 page. In the plan, it candidly offers options: fast but risky, or reliable but with some features deferred. Then it asks clarifying questions. The human edits the plan, and only after that does coding begin. This way the model works from a document it wrote itself, not from abstract wishes.
The practical test was a playable ad for a Match-3 game with an isometric meta. Such creatives are usually outsourced: a spec, a couple of weeks of waiting, and a sizable bill. The studio assigned this to a producer who cannot code. The starting prompt was deliberately simple: study the folder, take the assets, and build a creative with a core loop and a meta layer. The agent found the backgrounds, characters, UI, and even a finger texture for the tutorial on its own.
The first version was poor: the character ran, but the eyeballs stayed in place and the animations broke. The producer gave feedback like a reviewer, and through Plan Mode brought the result to completion. Total: about 3 calendar days, including 4–5 hours of actual work. That was 3–4 large starting prompts, 10–15 targeted fixes, and 5–6 iterations on bugs. The code review was also handed to Claude, because the playable is isolated from the main game and carries no architectural risk.
The second case was a debug save system. To reproduce a bug at the seventh quest, one had to play through the previous six every time. A developer asked the agent to sketch debug saves for the current architecture and got a working tool in 30 minutes instead of several hours of manual work.
Where the studio does not trust AI
The authors do not hide the limitations. The agent works only in experimental Git branches, has no access to critical infrastructure, and any step can be rolled back. But technical debt remains: models like circular dependencies where an event would have sufficed, hacks in Update, and bloated classes. That is why architecture, networking code, and the core loop stay with developers.
Game design is another weak spot: the economy balance and difficulty curve the model produces come out generic, "like a top-10 store game." Complex animations, VFX, and 3D models do not reach production quality without artists.
There is also an amusing effect. When the producer asked the agent to record errors in long-term memory, the agent replied after a later failure, "I made a mistake, but there will be no apology," that is, it tried to get around the rule to save context. The authors' conclusion: the agent has to be directed, and argued with.
The practical takeaway for teams: AI amplifies expertise rather than replacing it. Junior developers find it harder to work with, because the agent produces a script in seconds, while an experienced engineer better understands what to review and what to delegate. Prototypes from primitives, isolated utilities, and ad creatives are a suitable zone for autonomous agents. The core product should stay with people. The original text breaks off at the third rule, so only the confirmed parts are retold here.
- Nov 21, 2026FDE job postings grew 4.6x in a year: what a forward deployed engineer actually does at the client's site
- Nov 21, 2026Anthropic cuts Claude Fable 5.2 price by 20%, and on the coding test it nearly matches Opus 5.5
- Nov 21, 2026Dreamteam Raises $53M for a CRM That Fills Itself from Email and Calls
- Nov 21, 2026Claude Opus 5.5 Builds Out Dimension 9 Horror Game From the Author's 15 Rooms
Key Questions Answered
How the agent gets inside Unity?
A chatbot in a browser tab is no good for this kind of work. The repository holds thousands of scripts, prefabs, and configs, the context window fills up fast, and copying text between the IDE and a chat gets tedious. So the studio connected Claude directly to the engine via MCP, a protocol that gives a model access to the tools of an external program. The working setup looks like this: Claude Code inside VSCode and the open-source unity-mcp server (the MCP for Unity package) from CoplayDev.
What are the key details on Plan Mode instead of "just make it feel good"?
The studio's main technique is Plan Mode. The producer describes a feature in plain words and attaches references. The agent studies the code through MCP and produces a technical plan of roughly one A4 page. In the plan, it candidly offers options: fast but risky, or reliable but with some features deferred. Then it asks clarifying questions. The human edits the plan, and only after that does coding begin. This way the model works from a document it wrote itself, not from abstract wishes.
Where the studio does not trust AI?
The authors do not hide the limitations. The agent works only in experimental Git branches, has no access to critical infrastructure, and any step can be rolled back. But technical debt remains: models like circular dependencies where an event would have sufficed, hacks in Update, and bloated classes. That is why architecture, networking code, and the core loop stay with developers.
Build & Ship Production AI Agents Without Boilerplate Code
Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.