Part 2 of 4: The $50 Studio
What AI-native game development actually looks like when a professional runs it
Adam Clegg has been making games professionally since 2012. He worked as a level designer on PlanetSide 2, then as Senior Game Designer on H1Z1, where he led a team of 30 and grew the game from under 700 concurrent players to 12,700 concurrent and over a million monthly active users, with zero marketing spend. Creative Director at NantG. Most recently Studio Design Director at Liithos. He’s currently exploring the AI-native studio model he’s been building toward in public. Fifteen years of live service, multiplayer, online world design. He knows what a production pipeline looks like because he has run one.
He had been messing around with AI coding tools since 2022. Kept putting them down because, in his words, “the coding was never in a good enough spot to really talk to the AI and get something out of it that looked great, or played great, without a ton of extra effort hand editing it.”
That changed recently. And what he built once it changed is worth understanding in detail.
What He Built
Over roughly one month, Clegg built a vertical slice of a 2D Metroidvania, with combat areas modeled on Castlevania: Symphony of the Night. Not a demo reel. A functional prototype with connected rooms, enemy spawning and death handling, collision cleanup, damage rules, a miniboss encounter with an intro sequence and arena gates, a route-end boss, music switching for stage, miniboss, and end boss states, an SFX pass covering movement, attacks, pickups, shards, chests, enemies, and combat feedback, tool usage, shard pickups, HUD display, chests, item rewards, equipment flow, destructible torches, pickup drops, room button and gate logic with a timing obstacle, and a production map with boss room markers, gate indicators, and a legend.
The total cost in AI tools: $20 for Codex, $20 for Claude Pro, $10 for Suno. Fifty dollars.
Total focused work: about 60 hours over one month.
That number is worth sitting with. A vertical slice of a Metroidvania, built to a playable standard, with audio, with boss encounters, with a production map, for fifty dollars and sixty hours of one person’s time.
How He Did It
The workflow is the part that matters, and it’s not what most people picture when they hear “vibe coding.”
Clegg didn’t describe what he wanted and accept whatever came back. He built a small AI-assisted game studio around the project. Claude Code ran focused production lanes: gameplay programming, level design, audio direction, art direction, QA. Codex acted as technical lead and integration reviewer, checking architecture, validating diffs, protecting the main branch, accepting work only after review and testing. The structure mirrors a real studio org chart, just compressed.
He ran up to 8 agents in parallel, each handling a separate story, none stepping on the others. To get there, he and Codex did ground work first to set up the cluster, separating the agent stories so they could run without conflict. Then he fired the prompt, Claude assembled the agents, and they worked. He playtested the outputs, marked them as fixed or gave feedback, and the cycle continued.
Level design was the hardest part. The AI kept placing long diagonal ramps across entire rooms, blocking character movement, or stacking platforms too close together so the player couldn’t jump between them. Clegg worked through it in feedback loops: the agent learned that downward traversal required a platform at the end of a ramp, not just the ramp itself. It learned that upward traversal needed platforms spaced far enough apart to allow both clearance below and a jump above. Once the level design agent had enough knowledge to block out a full playtest area, Clegg moved on to other systems. The agent had, in his framing, leveled up.
The process he describes is not “tell AI to make game.” It’s closer to: scope a story, assign it to a specialized agent, validate the output against a QA checklist, give playtest feedback, iterate, and commit only after review. Structured story cards, memory files, relay packets, validation scripts. The AI made mistakes. The process caught them.
“It feels less like a one-off experiment now,” he wrote, “and more like a small production pipeline: creative direction, scoped implementation, automated validation, human playtesting, review, and iteration.”
What the Workflow Actually Requires
The skill the platform stage builds, and the agentic stage requires more of, is judgment.
Clegg’s 15 years of professional experience didn’t become irrelevant when he switched to AI tooling. They became load-bearing. He knew what good traversal felt like. He knew why a ramp placed diagonally across a room was wrong. He knew what an SFX pass should cover. He knew what a vertical slice needed to demonstrate. The AI handled a large portion of the implementation work. He handled the decisions about what to implement, whether it worked, and what to fix.
Owen Mahoney, former CEO of Nexon, describes this as AI operating as “power tools that help humans build.” In his framework, the first layer of AI in game development consists of tools that multiply human creativity and speed up production without replacing the human driving them. What Clegg built is a live demonstration of that layer operating at full capacity. The AI didn’t design the game. It built what a designer with 15 years of taste and experience told it to build, checked against acceptance criteria that designer wrote.
The important thing is that this workflow is not exclusive to people with 15 years of experience. The judgment required scales with what you’re trying to make. A creator on Rosebud who has spent 20 hours iterating on a platformer has developed real judgment about what works in that genre. That judgment is the asset. The agentic tools are what let you deploy it faster and at greater depth than manual implementation allows.
The Business Model Hiding in Plain Sight
Clegg posted something else worth flagging. After documenting the workflow, he described what he actually wants to do with it:
“A dream of mine right now would be to lead a nimble AI focused dev superteam of 5-6 people to create a game. Creative Design Director + AI. Tech Lead/Gameplay Engineer + AI. 2D/3D Artist + AI. Technical UI Artist + AI. Animator + AI. Audio + AI.”
His framing: “AI is used as the big machinery, coming in to dig the big holes, the cranes to lift things into place. People then use their hand crafted skills to add the paint, create all the finer details, place things into place that make it better. Nimble, focused, not stuck in meetings or production logistics. Just implementing, playtesting, and iterating toward a goal.”
This is a new studio model, and it’s not theoretical. The math works. A team of 6, each with a craft specialty and AI acceleration in their lane, building toward a $15 Steam game. No publisher required. No production management layer between the creative people and the work. The AI handles the machinery; the humans handle the taste.
Mahoney made the same point from the investor side, estimating that the right AI tooling could deliver what was once a $300 million AAA production for roughly $30 million. Clegg arrived at the same conclusion from the inside of a $50 prototype. The direction of travel is consistent.
What This Means if You’re Not Adam Clegg
The $50 vertical slice belongs to someone with 15 years of experience. That’s the honest version. The workflow Clegg built is not a beginner’s workflow, and pretending otherwise would be misleading.
But that raises the real question: where does the judgment come from, if not from 15 years in the industry?
Manik Bhat is one answer. He’s the solo developer behind Future Vibe Check, a Unity game currently wishlisted on Steam with Akupara Games signed as publisher, built over 2.5 years starting from zero coding experience. What’s notable about his approach is that he didn’t use AI to build the game whole-hog. He used AI to teach himself how to build it. The AI wasn’t the pipeline. It was the mentor. He asked it to explain what he didn’t understand, worked through problems iteratively, and accumulated the kind of practical game development knowledge that used to require a computer science degree or years of self-directed study.
That’s a different relationship with the tools than Clegg’s, and it points to a different rung on the same ladder. Clegg brought judgment to AI and used it to compress a team into one person. Manik used AI to acquire judgment he didn’t have, then built the game with it. The destination is the same. The starting point is different.
This matters because it means the path isn’t gated by prior experience. It’s gated by patience and the willingness to treat AI as a learning environment rather than a shortcut. The platform stage, covered in Part 1, is where you build taste through iteration. The AI-as-teacher stage is where you convert that taste into craft knowledge. The agentic pipeline stage, which is where Clegg operates, is where you deploy both. All three are available to anyone. None of them are instant.
That’s the path. It’s slower than a $50 month with 15 years behind you, and faster than it would have been without these tools. What it requires at every stage is the same thing: the decision to stop generating and start directing.
Part 3 looks at what happens when that logic reaches studio scale — and what it means that the answer there is more complicated than anyone expected.
In Part 3: How established studios are actually using AI across their pipelines, what it reveals about where the industry is heading, and why the gap between a 6-person AI studio and a 600-person AAA team is closing faster than either side expected.
Sources and further reading
Adam Clegg (arclegger): LinkedIn build-in-public posts, Apr–Jun 2026 https://www.linkedin.com/in/arclegger/
Adam Clegg — “A Journey with AI and my Son” (arclegger.itch.io) https://arclegger.itch.io/fishgame/devlog/1404637/a-journey-with-ai-and-my-son
Future Vibe Check (Steam wishlist) https://store.steampowered.com/app/3525890/Future_Vibe_Check/
daily.dev — “Vibe Coding in 2026: How It Works and When to Use It” https://daily.dev/blog/vibe-coding-how-ai-changing-developers-code/
