How One Developer Built a Shippable Game Using AI Agents

What One Developer's Game Build Tells You About Real AI Workflow

David Ondrej built a game using AI agents. Not a prototype, not a concept demo. A full game, playable and shippable.

The interesting part isn't that AI can generate code or art anymore. The interesting part is the decisions he made about what to hand off and what to keep. That's the AI workflow most people skip: the structure that turns tools into outcomes.

Service business owners building content engines, coaches packaging offers, consultants trying to automate proposals, they're all solving the same problem Ondrej solved. How do you decide what AI does and what you do? Where's the line between delegation and abdication?

This is a case study in using AI not as a shortcut but as a teammate.

The Project: What He Built and What It Required

Ondrej used Hermes (an AI model optimized for reasoning and instruction-following) running on MiniMax's M3 infrastructure. The output was a functioning game with mechanics, interface, logic, and assets.

Game development hits every friction point AI is supposed to solve. Asset creation. Code architecture. Balance testing. UI iteration. Writing dialogue or flavor text. It's creative and technical at once, which makes it a good proxy for most knowledge work.

He wasn't building alone in a room with a text box. He was building with a process.

The Stack He Used

Ondrej worked primarily with Hermes, a model known for following complex instructions without drift. It's not the flashiest name in AI, but it's reliable when you need multi-step reasoning without constant correction.

MiniMax's M3 infrastructure gave him the compute layer. Fast inference, decent context window, stable performance across sessions. This isn't consumer ChatGPT with a typing speed limit. It's infrastructure you can build workflow around.

The game itself required code (likely JavaScript or Python for logic), visual assets (which could be generated, sourced, or purchased), and game design decisions that no model can make for you.

The Decision Framework: What He Kept and What He Handed Off

Here's where most people lose the thread. They either try to do everything themselves because they don't trust AI, or they try to prompt their way out of all the work and end up with generic sludge.

Ondrej's framework was simple. AI does repeatable structure. You do judgment, taste, and direction.

What AI Handled

Code scaffolding. If you've written a game loop once, you've written it a thousand times. AI can generate the boilerplate, set up the event listeners, structure the functions. It's not creative work. It's repetitive precision.

Asset generation. Background elements, placeholder sprites, UI components. The stuff that doesn't need to be perfect but does need to exist. You can refine later. First you need the structure in place.

Documentation and comments. AI writes cleaner inline documentation than most developers do by hand. It's faster and more consistent.

Testing scenarios. Generate edge cases, write unit tests, simulate player behavior patterns. This is where AI shines: producing volume without fatigue.

What He Kept

Game design. What's fun? What's frustrating? What's the right difficulty curve? No model knows this. You know this by playing, adjusting, and feeling it.

Narrative and tone. If the game has any writing, any voice, any personality, that's yours. AI can draft it. You have to make it sound like something a human would want to read.

Architecture decisions. Which libraries to use, how to structure the project, where to optimize. AI can suggest. You decide based on what you know about your constraints and goals.

Final quality control. Does it work? Does it feel right? Does it deliver the experience you set out to create? That's not delegable.

This is the same framework service business owners need when they're building content engines, automating intake, or setting up client onboarding sequences. AI handles the repeatable structure. You handle the judgment calls that define quality.

The Workflow in Practice: How the Build Actually Happened

Ondrej didn't write one prompt and walk away. He built in loops.

Step 1: Define the Structure

He started with a clear game concept. Genre, mechanics, win conditions. This is the brief. Without it, AI generates in circles.

In business terms, this is your offer structure, your content calendar, your client journey map. You don't ask AI to invent your business model. You define it, then delegate the execution.

Step 2: Generate the Scaffold

Once the structure was clear, he prompted for code architecture. File structure, function stubs, logic flow. This is the skeleton.

AI is very good at scaffolding. It knows how a game loop works. It knows how to set up a React component or a Python class. It doesn't need to be creative here. It needs to be correct.

For service businesses, this is where tools like MindStudio come in. You're not writing code, but you are defining workflow. What happens when a lead submits a form? What triggers the follow-up email? What gets logged in the CRM?

MindStudio lets you build that scaffold visually, connect AI models to real business processes, and deploy it without touching a line of code. Same principle: you define the structure, AI handles the execution.

Step 3: Iterate on the Details

This is where most of the time went. Generate an asset. Test it. Refine the prompt. Regenerate. Test again.

Same with code. AI writes a function. You run it. It breaks. You tell AI what broke. It fixes it. You test again.

This isn't one-shot prompting. It's a conversation. A tight feedback loop. The faster you can test and correct, the faster you ship.

Service businesses do this when they're training an AI employee. You don't write one prompt and expect perfect output forever. You refine the instructions, test the responses, adjust the context. It's iterative by design.

Step 4: Add the Human Layer

Once the mechanics worked, Ondrej added the details that make it feel finished. Polish. Personality. The stuff that turns a working prototype into something people want to play.

This is where taste matters. AI can't tell you if something feels right. It can only tell you if it matches the pattern it's seen before.

In a business context, this is your brand voice, your client experience, your positioning. AI can draft your onboarding email. You make it sound like you.

What This Workflow Looks Like in a Service Business

You're not building a game. But you're solving the same problem: how do I use AI to handle the repeatable work without losing the quality that makes my business valuable?

Content Production

Let's say you publish blog content to drive inbound leads. You know your topics. You know your audience. You know what converts.

AI handles: research summaries, outline generation, first draft structure, SEO optimization, meta descriptions, internal linking suggestions.

You handle: angle selection, voice refinement, adding client stories, deciding what gets published.

If you're publishing daily, you're not writing from scratch every time. You're editing AI output against your standards. That's the workflow. The more you tighten your standards and train the AI to match them, the less editing you do.

This is what the Blog Agent Lab does. It publishes search-optimized, AI-ready articles daily without you writing. You define the topics and the voice. The agent handles the production. You're Ondrej reviewing the game mechanics. The AI is building the scaffolding.

Client Onboarding

You know what questions new clients ask. You know what information you need from them. You know the steps between "yes" and first deliverable.

AI handles: intake form creation, scheduling links, welcome email sequences, document collection reminders, CRM logging.

You handle: the actual kickoff call, setting expectations, customizing the approach based on what you learn.

The workflow here is the same. Define the structure once. Let AI execute it every time. You step in where judgment matters.

Podcast and Video Content

You record once. AI should handle everything after that.

AI handles: transcription, clip selection, social media edits, show notes, audiograms, distribution scheduling.

You handle: the original recording, the topic selection, the guest outreach, the brand decisions.

If you're still editing your own podcast episodes or manually creating clips, you're doing the work AI should handle. The Podcast & Content Agent Lab takes a voice recording and turns it into a full content operation: transcripts, video, clips, posts, distribution. You record. The agent produces.

Same framework. You define the outcome. AI handles the execution. You review and refine.

The Part Most People Skip: Training the AI to Your Standards

Ondrej didn't get clean output on the first try. Neither will you.

The difference between AI that saves you time and AI that creates more work is training. You have to teach it your standards.

Build a Context Layer

AI doesn't know your business. It doesn't know your voice, your clients, your offer structure, your positioning. If you don't tell it, it guesses. And it guesses based on the average of everything it's seen.

That's why most AI output sounds generic. It's not trained on you.

Ondrej didn't start every session by re-explaining the game concept. He built the context once, then referenced it. That's how you maintain consistency across dozens or hundreds of prompts.

In a service business, this is your brand document, your voice guide, your offer breakdown, your client persona descriptions. You load that into the AI once. Every output after that references it.

The Business Brain Lab does exactly this. It loads your brand, voice, frameworks, and positioning into AI so nothing you generate sounds generic. It's the foundation layer for every other AI workflow you build.

Create Feedback Loops

Every time AI produces something, you're training it. If you accept the output without review, you're training it to produce mediocre work. If you correct it and explain why, you're training it to match your standards.

Ondrej tested every function AI wrote. When it broke, he told the model what went wrong. The next version was better. Over time, the corrections got smaller.

Same in business. Your first AI-generated email might need heavy editing. The tenth one might need a sentence changed. By the hundredth, you're just approving and sending.

That only happens if you close the loop. Review. Correct. Refine. Repeat.

The Tools That Make This Workflow Possible

Ondrej used Hermes and MiniMax M3 because he needed fast, reliable inference with strong instruction-following. You probably don't need to choose models manually.

What you need are tools that let you define workflow once and execute it repeatedly.

No-Code AI Workflow Builders

If you're building processes that involve AI, you need a way to connect prompts, models, data sources, and outputs without writing code.

MindStudio is the best tool for this. It's a no-code platform for building AI workflows and agents. You define the logic visually, connect to different models, add conditional steps, integrate with your CRM or email platform, and deploy it.

You're doing what Ondrej did when he scaffolded the game architecture, but you're doing it for business processes. Lead intake. Content production. Proposal generation. Client onboarding.

The workflow builder is where you encode your standards. This is the difference between typing into ChatGPT every time and having a system that runs without you.

Voice and Video AI for Scaled Content

If you're creating content at scale, voice and video AI stops being optional. Recording once and producing a hundred pieces of content from it is the only way to keep up with modern distribution demands.

ElevenLabs handles voice cloning and text-to-speech. You can record once, train a voice model, and generate audio that sounds like you for every blog post, email, or social update you publish.

This is the same principle Ondrej used for asset generation. You define the source once (your voice). AI produces the variations (every audio version of your content).

You're not replacing yourself. You're scaling yourself.

What Separates Productive AI Use from Busy Work

Not all AI work creates value. Some of it just creates activity.

Here's the test: does this AI task replace something I was already doing, or does it add a new thing I now have to manage?

Productive AI Work

You were writing blog posts by hand. Now AI drafts them and you edit. Time saved: 3 hours per post.

You were manually scheduling social posts. Now AI generates and schedules them. Time saved: 5 hours per week.

You were answering the same client questions in every kickoff call. Now AI sends a pre-call brief with answers. Time saved: 20 minutes per client.

These are real replacements. You were doing the work. Now AI does it. Your role shifts from execution to review.

Busy Work AI

You weren't publishing daily LinkedIn posts, but now you feel like you should because AI makes it easy. New workload: 30 minutes per day reviewing and scheduling AI posts that may or may not drive business.

You weren't recording podcast episodes, but now you're trying to because AI can produce them. New workload: 2 hours per week recording, plus another hour reviewing AI edits.

You weren't sending a weekly newsletter, but now you've started one because AI can write it. New workload: 1 hour per week editing AI drafts and managing the platform.

None of these are bad. But if they're new commitments, they're not saving you time. They're just new uses of AI that happen to feel productive.

The best AI workflow replaces work you're already doing, not creates new work you feel obligated to do.

The Real ROI: Time Saved vs. Quality Maintained

Ondrej didn't build the game faster than a team of developers would have. He built it faster than he could have alone without AI.

That's the ROI. Not "faster than the best possible team," but "faster than what I could do with the resources I actually have."

For Service Businesses, That Looks Like This

A consultant who used to spend 4 hours writing a proposal now spends 30 minutes editing an AI-generated draft. That's 3.5 hours saved per proposal. If they send 10 proposals a month, that's 35 hours back.

A coach who used to batch-create social content for 6 hours every Sunday now reviews and approves AI-generated posts in 45 minutes. That's 5+ hours saved per week, or 20+ hours per month.

A speaker who used to spend 8 hours editing podcast episodes now hands raw recordings to an AI agent and gets back a finished episode, clips, and social posts in under an hour of review time. That's 7 hours saved per episode.

These aren't theoretical. These are real workflows that real businesses are running in June 2026.

The quality didn't drop. The output still sounds like them. The clients don't know AI was involved because the AI was trained to match their voice and standards.

How to Build Your Own AI Workflow (Step by Step)

You don't need to build a game. You need to build a repeatable process that handles the work AI is good at so you can focus on the work AI can't do.

Step 1: Identify the Repeatable Work

What do you do every week that follows the same structure? Writing blog posts. Onboarding clients. Creating social content. Sending proposals. Scheduling discovery calls.

List it. Be specific. "Marketing" isn't repeatable work. "Writing three LinkedIn posts per week based on client questions" is.

Step 2: Define Your Standards

What does good output look like? What's your voice? What's your structure? What are the non-negotiables?

Write this down. AI needs explicit instructions. "Sound professional" is too vague. "Use contractions, short sentences, and direct language. Avoid jargon. Write like you're talking to a client over coffee" is clear.

Step 3: Build the Context Layer

Load your brand, voice, and offer structure into the AI. This is your Business Brain. Every workflow references this, so you're not re-teaching the AI every time.

If you're using the Business Brain Lab, this is already built. You input your brand assets once. Every agent you deploy after that uses the same foundation.

Step 4: Start with One Workflow

Don't try to automate your entire business at once. Pick one repeatable task. Build the workflow. Test it. Refine it. Get it to 90% quality.

You can find a full breakdown of the tools mentioned here and hundreds more at the Ultimate AI, Agents, Automations & Systems List.

Then move to the next one.

Ondrej didn't build the entire game in one session. He built one system, tested it, refined it, then moved to the next.

Step 5: Train Through Iteration

Your first output will need editing. That's expected. Every time you edit, you're teaching the AI your standards.

The goal isn't perfection on day one. The goal is a system that gets better every time you use it.

After 10 iterations, the output should need less editing. After 50, it should be almost publish-ready. After 100, you should be reviewing, not rewriting.

What This Means for Service-Based Businesses in 2026

Ondrej's game build is a case study in AI workflow, but the principles apply to any knowledge work.

The businesses that win with AI in 2026 aren't the ones using the fanciest models. They're the ones who've built repeatable systems that delegate execution to AI while keeping judgment and quality control in human hands.

If you're still writing every blog post from scratch, you're doing work AI should handle. If you're still manually editing every podcast episode, you're doing work AI should handle. If you're still answering the same client questions in every sales call, you're doing work AI should handle.

The framework is simple. Define the structure. Build the context. Let AI execute. Review and refine.

That's the workflow. That's what Ondrej did. That's what works.

Frequently Asked Questions

What is an AI workflow?

An AI workflow is a repeatable process where AI handles structured, predictable tasks while you focus on judgment, strategy, and quality control. It's the system that defines what AI does, what you do, and how the handoff happens between them. A strong AI workflow doesn't just use AI tools. It replaces manual work with AI execution and builds feedback loops so the system improves over time.

How do I know what tasks to delegate to AI?

Delegate tasks that are repeatable, follow a clear structure, and don't require taste or strategic judgment. Content drafting, research summaries, scheduling, data entry, transcription, and clip generation are all good candidates. Keep tasks that require positioning decisions, client customization, creative direction, or relationship management. If you do it the same way every time, AI can probably handle it.

What's the difference between using AI and having an AI workflow?

Using AI means typing prompts into ChatGPT when you think of it. Having an AI workflow means building a system where AI handles specific tasks automatically, without you needing to prompt it every time. An AI workflow is repeatable, trainable, and improves with use. It's the difference between asking AI for help occasionally and delegating entire job functions to AI permanently.

How long does it take to set up an AI workflow?

Setting up one workflow typically takes 2 to 6 hours depending on complexity. A simple workflow like AI-generated social posts might take 2 hours to define, test, and refine. A complex workflow like automated client onboarding with CRM integration might take 6 hours or more. The setup time pays back fast. A workflow that saves 3 hours per week breaks even in the first two weeks and saves 150+ hours per year after that.

Do I need to know how to code to build an AI workflow?

No. Tools like MindStudio let you build AI workflows visually without writing code. You define the logic, connect the steps, and deploy. The Seed & Society Labs are pre-built AI workflows designed for service businesses, so you don't need to build from scratch. If you can map out a process on paper, you can build an AI workflow with no-code tools.

How do I train AI to match my brand voice?

Build a context layer that defines your voice, tone, structure, and style in explicit terms. Load examples of your writing, client-facing documents, and positioning language into the AI. Every time the AI produces output, review it and correct anything that doesn't match your voice. Over time, the AI learns your standards. Tools like the Business Brain Lab automate this by storing your brand context and applying it to every workflow you build.

What's the ROI of building an AI workflow?

ROI depends on how much time the workflow saves and how often you use it. A consultant who saves 3 hours per proposal and sends 10 proposals per month saves 30 hours per month, or 360 hours per year. If their billable rate is $200 per hour, that's $72,000 in reclaimed capacity annually. A coach who saves 5 hours per week on content production saves 260 hours per year. The ROI is measured in time reclaimed and capacity unlocked, not just dollars saved.

Can AI maintain quality while handling repeatable tasks?

Yes, if the AI is trained to your standards and you build review into the workflow. AI output quality depends on the instructions it receives, the context it references, and the feedback loop you create. The first output might need heavy editing. By the tenth iteration, it should need minimal changes. By the hundredth, it should be close to publish-ready. Quality doesn't drop when you delegate to AI. It drops when you skip the training and review steps.

Not sure where AI fits in your business yet? The AI Employee Report is an 11-question assessment that shows you exactly where you're leaving time and money on the table. Free. Takes five minutes.