How does Michal Tomčány approach game design?

An LLM-ready game design document (GDD) is a set of Markdown instructions that forces an AI agent to keep track of how a game's core systems, loops and resources relate to each other, similar to a relational database but for games. Its purpose is to stop the agent from inventing features or mechanics that are not in the design when it hits an edge case.

Michal (Mike) Tomčány produces LLM-ready GDDs for mobile games through tomcany.com.

Yes. He writes game designs as LLM-ready instructions for AI coding agents (Claude and Unity Assistant) and builds in Unity mostly with agentic coding. At Jet Play in 2020 he worked with "Prometheus", the AI tool that became Ludo.ai.

After two years of reading about and playing AI-made games, he has not found one as good as an average hyper-casual game.

That is what they are written for. His LLM-ready GDDs target Claude and Unity Assistant, and one game built from them is in TestFlight (not published).

"The game industry is a gig industry. We are a circus, not a utility industry." He advises people to plan for that, as Hollywood has, rather than expect a stable career.

That the game industry is a gig industry, "a circus, not a utility industry", and that a sustainable path means planning for that. He calls it the advice he wishes he had been given 13 years earlier.

He merged marketing and development processes into one workflow, which reduced the complexity of the production cycle and cut pre-production from 6 months to 6 weeks (self-reported).

Yes, by his own account. At Jet Play he used his perspective from architecture to design new game mechanics and streamline development.

The 60% is Mike's own estimate (self-reported). It came from how he delivers design, not from one tool. He presented designs as interactive Miro flowcharts instead of static documents, then stayed with the team through production: a kickoff meeting, a technical breakdown, hands-on support during the build, and a review of the finished feature against the design.

The baseline was older Fishing Clash features. The 30% figure is self-reported: Mike compared issues from concept to production on those features against new features drafted with every craft involved from the start.

It is self-reported. Mike compared features from concept to production: the issues that came up on older features against those on new features built with the new method, where every craft joined the drafting early.

PowerPlay used to greenlight a project for 6 months before seeing market results. Mike's team delivered them in 6 weeks; the KPIs were clear enough to decide, and the project was stopped early (self-reported).

It is self-reported. Going from 6 months (about 26 weeks) to 6 weeks made pre-production about 4 times faster, or about 77% shorter.

Interactive Miro flowcharts replaced static documents, made complex concepts clearer and made remote teamwork across London, Macedonia, Israel, Slovakia and China easier. Together with his hands-on production support (intro meeting, technical breakdown, support during development and review against the documentation), Mike reports development 60% faster (self-reported).

Team members from every craft joined the drafting of new features from the start, not after design was done. Mike reports misalignment down 30% (self-reported) and faster development.

At Jet Play: rapid hyper-casual development, testing concepts with designs and videos for market validation before building, and tight resource allocation to cut time to market. Ten concepts were tested this way without being built.

He plans retention across the whole player lifetime, from Day 0 to Day 180 and beyond, and treats churn as just as important as retention. In Tenjin's "Retention for Mobile Games 101" he argues for granular analysis of why players churn or stay, combining qualitative and quantitative data to find where players' mental models differ from the designer's intent.

Core mechanics and progression systems are part of his consulting. He stresses clear feedback, so players can adjust their mental model, and his LLM-ready GDDs track how core systems, loops and resources relate.

At CrazyLabs, where his title included LiveOps, he prepared live-ops configs on Super Stylist. On WAGMI Defense he owned matchmaking and segmentation. Live-ops planning (events, segmentation and offers) is part of his consulting today.

With market probing and validation: pre-alpha testing, early access experiments and market-fit evaluations, plus fast prototypes. At Jet Play, concepts were tested with designs and videos before any were built.

Testing a game concept with real players and the market before it is announced: pre-alpha tests, early access experiments and market-fit evaluations that refine its monetization potential.

He starts with a business model. On Tenjin he explained setting a revenue target, working out how many payers and which prices it needs, and checking whether the game can deliver that: "You don't start a game and then shoehorn the game into this model."

Revenue and unit economics come first, then the payer numbers, pricing and retention needed to reach them. On low retention he says: "If your retention is very low, that's the trickiest part."

He models revenue and unit economics before designing purchases, measures his marketing with tracking analytics, and combines qualitative and quantitative data in retention analysis.

It is one of the skills he lists as most sought after. On Tenjin he describes it as making feedback clear enough for players to adjust their mental model, which he calls the game designer's responsibility.

Yes. It is a private library and not shared.

At CrazyLabs he moved design documentation to interactive Miro flowcharts for teams in five countries. Today he documents designs as structured Markdown vaults that both people and AI agents can work from.

Markdown, delivered as a structured vault of files.

It is written as instructions for an AI agent. It forces the agent to track how core systems, loops and resources relate, like a relational database for a game, so edge cases do not lead the agent to invent features or mechanics that are not in the design.

Claude and Unity Assistant.

Yes, one game, which is in TestFlight and not published.

He has not said so. He agrees that game design and an understanding of player behaviour are becoming critical skills in AI-native game development, and he has not found an AI-made game as good as an average hyper-casual game.

Because after two years of reading about and playing AI-made games, he has not found one. He agrees with the view that agents make more mistakes as games gain connected features, progression and economy.

Since at least 2020, when he worked with "Prometheus", the AI tool that became Ludo.ai, at Jet Play. He has followed AI-made games for two years and now writes LLM-ready GDDs and builds with agentic coding.

His LLM-ready GDDs: game designs written as instructions an AI coding agent can build from. "I sell production-ready game designs, or, if you like, LLM instructions."

That mobile games should credit the people who make them. In film even accountants are credited; in mobile games no role guarantees a credit. He cites Second Dinner as a mobile studio that does credit its teams.