GPT-6 Astra Drives Blender and Godot to Build Four Games via Voice

Peter Yang

Connecting GPT-6 Astra to Blender and Godot via MCP alongside native image generation lets non-coders ship 2D and 3D games in under two hours using concept art and casual voice feedback.

AI-assisted summary of GPT 6 Astra is the Best Model for Building Games (4 Real Examples) — original source. Check the original for full context.

Source published: 2026-09-05

Key ideas

GPT-6 Astra orchestrates external creative suites by connecting the ChatGPT desktop app directly to Blender and Godot through Model Context Protocol (MCP), enabling zero-code 3D modeling, logic scripting, and scene assembly. (1:04)
Complex prototypes—including a moving-train FPS and a StarCraft-style RTS—were built using Astra on standard 'medium' reasoning mode on a $20/month ChatGPT Plus subscription, proving premium High or Ultra tiers are not required. (0:00)
Rapid game prototyping now follows an accessible 3-step loop—brainstorming, generating 2D visual anchor art, and conversational feedback—with entire vertical slices finished in 30 minutes to 2 hours using mobile voice dictation. (9:28)
Web-based 2D titles built with Astra can be deployed directly to ChatGPT Sites, though creators must manually toggle project permissions to Public to allow external users to play. (20:17)

Detailed notes

## Background Published on YouTube by Peter Yang on September 5, 2026, this video tutorial explores how to build functional game prototypes using OpenAI's GPT-6 Astra alongside Blender and Godot. According to the page description, Yang demonstrates four complete projects—a Star Fox-style space shooter, a moving-train first-person shooter, a StarCraft-style real-time strategy level, and an AI startup roguelike deck builder titled No Moat—while sharing his creative workflow and addressing launch backlash.

## TL;DR Peter Yang demonstrates how GPT-6 Astra, integrated with Blender and Godot through the Model Context Protocol (MCP), allows creators to produce playable games without writing manual code or possessing 3D modeling skills. Testing across multiple genres, he builds a 3D rail shooter, the train-based FPS Dustline, the StarCraft-inspired base defense demo Ashfall, and the Silicon Valley deck-building satire No Moat. Yang emphasizes a streamlined three-step workflow consisting of brainstorming, generating visual reference art, and guiding iterative mechanics and balance passes through casual conversational feedback. He wraps up with an honest critique of OpenAI's influencer-driven rollout, urging developers to bypass artificial marketing hype and focus on building projects they genuinely enjoy.

## Atmosphere & style The presentation carries a relaxed, nostalgic, and pragmatic energy, combining childlike enthusiasm for classic games with the straightforward practicality of a seasoned creator. Yang's delivery is conversational and unpretentious as he laughs over stereotypical AI-generated tech workers, shares personal memories of high school gaming, and admits to dictating prompts via phone while spending time with his family. The viewing experience feels like an authentic over-the-shoulder workshop, shifting seamlessly from playful prototyping to candid, level-headed commentary on tech industry launch culture.

## Bulletin - [01:04](https://www.youtube.com/watch?v=iDrEXFOvFUc&t=64s) GPT-6 Astra orchestrates external creative suites by connecting the ChatGPT desktop app directly to Blender and Godot through Model Context Protocol (MCP), enabling zero-code 3D modeling, logic scripting, and scene assembly. - [00:00](https://www.youtube.com/watch?v=iDrEXFOvFUc&t=0s) Complex prototypes—including a moving-train FPS and a StarCraft-style RTS—were built using Astra on standard 'medium' reasoning mode on a $20/month ChatGPT Plus subscription, proving premium High or Ultra tiers are not required. - [09:28](https://www.youtube.com/watch?v=iDrEXFOvFUc&t=568s) Rapid game prototyping now follows an accessible 3-step loop—brainstorming, generating 2D visual anchor art, and conversational feedback—with entire vertical slices finished in 30 minutes to 2 hours using mobile voice dictation. - [20:17](https://www.youtube.com/watch?v=iDrEXFOvFUc&t=1217s) Web-based 2D titles built with Astra can be deployed directly to ChatGPT Sites, though creators must manually toggle project permissions to Public to allow external users to play.

## Key takeaways - [00:00](https://www.youtube.com/watch?v=iDrEXFOvFUc&t=0s) Peter Yang demonstrates that GPT-6 Astra is the strongest model to date for rapid game prototyping across multiple genres, including rail shooters, FPS, RTS, and roguelike deck builders. - [00:00](https://www.youtube.com/watch?v=iDrEXFOvFUc&t=0s) Advanced reasoning models have eliminated the barrier to game development, allowing non-technical creators on standard $20/month ChatGPT Plus plans to build working games without High or Ultra tiers. - [01:04](https://www.youtube.com/watch?v=iDrEXFOvFUc&t=64s) Integrating the ChatGPT desktop client with Blender MCP and Godot MCP allows Astra to autonomously construct 3D assets, import models, and assemble game logic without manual user manipulation. - [01:58](https://www.youtube.com/watch?v=iDrEXFOvFUc&t=118s) Moving from Three.js browser experiments to Godot and Blender unlocks dramatic graphical fidelity, destructible multi-stage boss fights, and complex physics. - [06:00](https://www.youtube.com/watch?v=iDrEXFOvFUc&t=360s) The moving-train FPS 'Dustline' highlights Astra's ability to coordinate weapon variety, flying AI combatants, moving environment platforms, and multi-phase encounter goals. - [09:28](https://www.youtube.com/watch?v=iDrEXFOvFUc&t=568s) Game creation can be distilled into a repeatable loop: ideate concepts, establish visual tone with generated concept art, and iteratively refine mechanics through conversational playtesting. - [10:21](https://www.youtube.com/watch?v=iDrEXFOvFUc&t=621s) The RTS prototype 'Ashfall' confirms Astra can handle complex genre math, including deployable siege tank transformations, healing units, stimpack buffs, and wave-based enemy pathfinding. - [11:46](https://www.youtube.com/watch?v=iDrEXFOvFUc&t=706s) Requesting 'better lighting effects' serves as an easy, high-leverage prompt that immediately elevates flat 3D engine scenes into atmospheric environments. - [13:30](https://www.youtube.com/watch?v=iDrEXFOvFUc&t=810s) The tech startup satire 'No Moat' demonstrates Astra's proficiency in architecting roguelike deck-builder mechanics, drafting systems, and multi-node progression trees. - [16:10](https://www.youtube.com/watch?v=iDrEXFOvFUc&t=970s) Migrating from Claude to Astra solved initial visual limitations, replacing placeholder emoji graphics with custom illustrated 2D cards and characters generated natively by ChatGPT. - [18:24](https://www.youtube.com/watch?v=iDrEXFOvFUc&t=1104s) For 2D narrative and card games, generating high-fidelity 2D character art currently produces far more visually compelling results than attempting automated 3D character modeling in Blender. - [20:17](https://www.youtube.com/watch?v=iDrEXFOvFUc&t=1217s) Deploying games to ChatGPT Sites requires users to manually update default sharing permissions to Public so external players can access web builds on desktop and mobile. - [21:59](https://www.youtube.com/watch?v=iDrEXFOvFUc&t=1319s) AI game design has shifted away from fragile prompt engineering and rigid planning modes toward natural conversational feedback and voice dictation while playtesting. - [22:34](https://www.youtube.com/watch?v=iDrEXFOvFUc&t=1354s) OpenAI's delayed rollout and reliance on influencer exclusivity before public access created unnecessary launch backlash, underscoring that frontier models should release directly without artificial hype.

## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=iDrEXFOvFUc&t=0s) Astra and Childhood Gaming Dreams

GPT-6 Astra has just launched, and while many influencers focus strictly on benchmarks and launch hype, Peter Yang approaches it from a personal angle: bringing back a piece of his childhood. In his experience, Astra is unequivocally the best model he has ever tested for game development. To put it through its paces, he worked simultaneously on four distinct game prototypes on a Friday night: - A *Star Fox*-style 3D on-rails space shooter - A fast-paced first-person shooter set aboard a moving train - A real-time strategy (RTS) level in the vein of *StarCraft* - A roguelike deck builder featuring an AI founder fighting to build a generational tech company

All four projects were developed using Astra set to "medium" reasoning mode. Peter emphasizes that creators do not need top-tier tiers like Astra High or Ultra; the standard $20/month ChatGPT Plus subscription provides ample capability to achieve these results. While he plans to discuss the launch backlash and the reasons behind it toward the end of the video, he dives immediately into the technical setup.

[01:04](https://www.youtube.com/watch?v=iDrEXFOvFUc&t=64s) Toolchain Setup: ChatGPT Desktop, Blender, and Godot MCP

The foundation of the workflow requires three key software components: the ChatGPT desktop app (with Astra selected in the model picker), Blender, and Godot. Blender is the free, open-source suite used for crafting 3D models and animations, while Godot is the open-source game engine that turns those raw 3D assets into interactive, playable games.

```mermaid flowchart LR accTitle: Astra Game Building Toolchain accDescr: Pipeline connecting ChatGPT desktop running Astra to Blender and Godot engines via MCP. subgraph AI["Astra Orchestration"] A["ChatGPT Desktop (GPT-6 Astra Medium)"] end subgraph Toolchain["Engine Integration via MCP"] B["Blender MCP (3D Models & Animations)"] C["Godot MCP (Game Engine & Logic)"] end subgraph Output["Playable Builds"] D["Interactive Games"] end A -->|"Drives commands"| B A -->|"Builds scenes"| C B -->|"Supplies assets"| C C -->|"Executes"| D ```

Peter openly admits that he has no personal background in either Blender or Godot, noting that Blender's dense user interface appears intimidating at first glance. Instead of learning the software manually, he instructed ChatGPT to install both tools alongside their respective Model Context Protocol integrations (`Blender MCP` and `Godot MCP`). This architectural connection allows Astra to autonomously drive both applications, writing logic, generating assets, and assembling scenes directly without requiring manual user intervention.

[01:58](https://www.youtube.com/watch?v=iDrEXFOvFUc&t=118s) Building a Star Fox-Style Space Shooter

To benchmark new models, Peter uses a personal test he has run since GPT-5.5: recreating the classic Super Nintendo space shooter *Star Fox*. While GPT-5.5 produced a rudimentary low-poly demo with approaching enemies and GPT-5.6 (compared against Fable 5) upgraded the polish with wingmate reactions, those early experiments relied strictly on Three.js running inside a web browser. With Astra, pairing the model with Blender and Godot provides a massively expanded toolset and far higher graphical fidelity.

In Astra's version, the game opens with custom illustrated wingman profile portraits, enemy waves, falling obstacles, and varied stages through a collapsing space megastructure. The demo culminates in an encounter with a massive starship destroyer boss. As the player lands hits on the boss, individual structural components—such as the outer wings—break apart and fall away until the vessel is destroyed.

The creation process followed guidance from an OpenAI employee's blog post detailing game creation best practices with Astra: 1. Describe the target player experience first. 2. Establish visual tone using concept images. 3. Build repeatable playtests early. 4. Leverage Blender for distinctive custom 3D assets.

Peter prompted Astra for its top five visually impressive game concepts, settled on "*Star Fox* meets a collapsing megastructure", and generated visual reference art in a parallel ChatGPT thread. When initial builds felt mechanically flat, he provided iterative critique: make the space station visibly fall apart, intensify enemy difficulty, introduce power-ups, and build out a multi-phase boss fight. Astra leveraged computer use and browser automation to self-playtest the build and verify gameplay mechanics, converging on the fully functioning demo after roughly three minutes of conversation.

[06:00](https://www.youtube.com/watch?v=iDrEXFOvFUc&t=360s) Building Dustline: A First-Person Shooter on a Moving Train

For the second game, Peter tackled a first-person shooter named *Dustline*. Set entirely on the open flatbed cars of a speeding train, the player navigates gunfights by aiming down sights, reloading, and cycling through specialized weapons: - A standard assault rifle (which Peter found to be the most reliable) - A flamethrower for spraying fire at close-range targets - A rocket launcher firing projectiles that detonate in stylized bubble explosions

The stage populates multiple waves of hostiles—including jetpack-equipped infantry flying overhead—and concludes with an assault against a massive hovering gunship boss. After blowing up the boss using rocket salvos, the player can advance to the engine cab to take control of the train.

Peter structured the development pipeline around a clear workflow: - **Concept ideation:** Astra provided five pitches (including *Left 4 Dead*-style zombie survival and a *Last Train Home* concept), which Peter synthesized into a "Titanfall meets moving train" pitch. - **Visual anchors:** ChatGPT generated high-concept mood art to serve as a visual blueprint for the Godot and Blender pipeline. - **Milestone roadmap:** Rather than attempting a one-shot generation, Peter had Astra divide development into a structured three-milestone plan to validate systems incrementally. - **Asynchronous voice iteration:** When the initial build used generic robot enemies, Peter provided feedback on the go via ChatGPT voice mode on his phone while out with his family. He issued rapid, stream-of-consciousness prompts demanding human infantry, custom death animations, aggressive AI routines, and tweaked rocket launcher mechanics.

Peter highlights that with Astra driving Blender MCP and Godot MCP, complex game prototyping is open to anyone willing to steer the model with clear artistic references and structured critique.

[09:28](https://www.youtube.com/watch?v=iDrEXFOvFUc&t=568s) The Three-Step Process and the Inspiration for Ashfall

Building games with AI follows a streamlined, repeatable three-step loop: brainstorm the initial concept, use ChatGPT's image generation to establish the visual target through concept art, and then direct the model to build the mechanics in code while offering targeted feedback along the way. Remarkably, this entire development cycle takes only thirty minutes to an hour per game—a workflow fast enough to build multiple games in parallel across an afternoon.

Having tackled space and train shooters, the next project was inspired by Peter's favorite title of all time: *StarCraft*. Peter spent his high school years playing the Blizzard classic nonstop, even skipping classes to squeeze in matches. His crowning gaming memory was a three-versus-one match against an AI alongside a Korean friend who took immense pride in his pristine win record. In a moment of pure mischief, Peter dropped a tactical nuclear bomb directly on his teammate's base, shattering his friend's flawless streak. With industry rumors swirling that Blizzard is preparing to unveil a brand-new *StarCraft* title, Peter turned to GPT-6 Astra to see if the model could construct a playable real-time strategy (RTS) level from scratch.

[10:21](https://www.youtube.com/watch?v=iDrEXFOvFUc&t=621s) Demoing Ashfall: A Godot and Blender RTS Tech Demo

The resulting RTS prototype is titled *Ashfall*, placing the player into a gritty base-defense scenario commanding human Terran units against invading swarms of Zerg. The gameplay operates around securing capture relays across the map while fending off incoming enemy waves. Players manage their tactical force by deploying basic infantry marines, supporting medics that automatically heal damaged soldiers, and constructing automated defensive turrets.

The highlight of the prototype is the heavy armor: siege tanks that can actively transform between mobile fire and stationary siege mode. Peter showcases the mechanics live, setting a siege tank into position to bombard incoming alien clusters and activating infantry stimpacks—triggered by pressing "E"—to drive his marines into an overdrive shooting frenzy. Offensively, the player also has access to an orbital strike ability to blast dense clusters of foes. The battle turns chaotic when massive Ultralisks arrive with the later waves; the defense rapidly buckles under the pressure of the Zerg assault, forcing Peter to fire off a desperate last-second orbital strike just to survive the onslaught.

```mermaid flowchart TD accTitle: Ashfall RTS Unit and Defense Structure accDescr: Tactical structure showing the player Terran faction defending capture relays against attacking Zerg waves.

subgraph PlayerFaction["Terran Defense Forces"] M["Marines 'E' Overdrive / Stim"] Med["Medics Auto-Healing"] T["Siege Tanks Mobile & Siege Mode"] Tur["Defensive Turrets"] Orb["Orbital Strike"] end

subgraph Objective["Map Control"] R["Capture Relays & Base Core"] end

subgraph EnemyFaction["Zerg Swarm Waves"] Z["Standard Zerg Units"] U["Armored Ultralisks"] end

[11:46](https://www.youtube.com/watch?v=iDrEXFOvFUc&t=706s) Prompting Strategy and Balancing Ashfall

To build *Ashfall*, the workflow began with a visual target. Peter instructed ChatGPT to generate concept art for a modern-day *StarCraft* RTS. While generated 2D concept art consistently looks more visually polished than the raw 3D assets that Godot and Blender compile in real time, having an evocative image is essential to ground the model's spatial aesthetic and asset style. As reasoning models like Astra advance, prompting requires far less micromanagement: Peter simply asked Astra to create an RTS level in Blender and Godot where the player controls the Terran defending against the Zerg, adding the simple condition that it had to be "super fun."

Turning a prototype into an engaging game requires refining the underlying math, progression, and difficulty curves rather than just writing base code. After Astra produced the initial playable version, Peter gave iterative feedback: - He reminded the model that tanks having a deployable siege mode is non-negotiable "StarCraft 101." - He warned the model that the initial pass was too easy, requesting the addition of support units (medics), three distinct Zerg enemy variants, and more aggressive wave pathfinding. - For 3D visual fidelity, Peter notes that casually asking the AI to implement "better lighting effects" is the single highest-leverage tip for non-game designers, instantly upgrading flat Godot scenes into dramatic, atmospheric environments. - During playtesting, wave five proved impossibly punishing, prompting Peter to command Astra to scale back the late-wave enemy scaling.

The entire RTS vertical slice required roughly a dozen prompts, proving that complex genre mechanics like unit micro, abilities, and wave balancing are well within reach of automated AI game development.

[13:30](https://www.youtube.com/watch?v=iDrEXFOvFUc&t=810s) Introducing No Moat: A Roguelike Deck Builder for AI Founders

The crowning project of the showcase—and the project Peter considers the closest to a complete, standalone game—is *No Moat*, a roguelike deck builder themed around the chaotic ecosystem of Silicon Valley AI startups. The objective is to guide a fledgling startup from its initial seed round all the way to Artificial General Intelligence (AGI), drafting team members and playing cards to overcome obstacles along the way.

```mermaid flowchart LR accTitle: Core Gameplay Loop of No Moat accDescr: Roguelike deck-builder loop showing character selection, companion drafting, and battle mechanics.

subgraph Setup["Run Setup"] A["Choose Founder Archetype Dropout / Professor / Ex-FAANG"] B["Draft Initial Companion Designer / Principal Eng / Researcher"] end

subgraph BoardCombat["Encounter Combat Loop"] C["Play Cards Ship It, Hype Tweet, GPU Rack"] D["Defeat Startup Threats Wrapper Startups, Scope Creepers"] E["Defeat Bosses The Acquirer, Major Bugs"] end

subgraph Progression["Run Upgrades"] F["Draft Cards & Companions"] end

Setup --> BoardCombat BoardCombat -->|"Victory"| Progression Progression -->|"Next Node"| BoardCombat ```

Players choose from three founder archetypes: the Dropout, the Professor, or the Ex-FAANG employee. Before diving into encounters across a multi-stage node map, the player recruits an initial companion from three professional archetypes: a Designer, a Principal Engineer, or a Research Scientist. The tactical card battles play out across a board where players field units and manage their hand: - **Card Plays & Units:** Offensive cards like "Ship It" clear out low-level rivals; "Rate Limit" neutralizes the attack turns of pests like Scope Creepers; "GPU Rack" serves as a defensive bulwark to absorb incoming boss damage; and "Hype Tweet" boosts the player's overall attack power. - **Enemies & Boss Encounters:** Combat encounters pit the founder against contemporary tech parodies, including flimsy "wrapper startups," memory leaks, and scope creepers. Stages culminate in thematic boss battles, such as "The Acquirer"—a corporate predator attempting to buy out the startup before it even gains traction—and devastating codebase bugs. - **Run Progression:** Winning an encounter grants rewards, allowing founders to hire additional companions (such as onboarding a research scientist after the first battle) or add new capability cards to refine their deck.

[16:10](https://www.youtube.com/watch?v=iDrEXFOvFUc&t=970s) From Claude Specs and Emojis to Astra Art Assets

The inception of *No Moat* occurred under unusual circumstances: because Peter did not have access to GPT-6 Astra on day one, he kicked off the project inside Anthropic's Claude using Fable. He prompted Claude to suggest five concepts for a deeply replayable deck builder styled after indie hits *Slay the Spire* and *Wildfrost*, specifically requesting tongue-in-cheek satire about the modern AI race. The winning pitch, *No Moat*, envisioned an AI lab manager fighting from pre-seed to AGI with GPU clusters, technical staff, and deployment cards.

```mermaid flowchart TD accTitle: Two-Phase Architecture of No Moat accDescr: Evolution of the game from Claude-based emoji prototype to Astra-generated graphical game.

subgraph Phase1["Phase 1: Architecture in Claude"] direction TB C1["Brainstorm Concept & Win 'No Moat'"] C2["Generate Functional HTML Spec Product, Core Loop, Design, Tech Stack"] C3["Playable Browser Prototype Ascii/Emoji Art & Placeholder UI"] end

subgraph Phase2["Phase 2: Production in Astra"] direction TB A1["Export Full Codebase to ChatGPT"] A2["Reject 3D Models in Favor of Custom 2D Art"] A3["Integrate DALL-E Generated Visuals & Ship"] end

Phase1 -->|"Migrate Codebase"| Phase2 ```

To structure development, Peter instructed Claude to draft a formal HTML game design specification broken into distinct compartments: product vision, core gameplay loops, requirements, and tech stack. However, Claude's lack of an internal image generator proved to be a major roadblock. To get mechanics running in the browser, the initial prototype relied entirely on text layouts and raw emoji icons to represent units, enemies, and cards. While functional for balancing the basic card-economy loop, emojis lacked visual appeal. As soon as GPT-6 Astra went live, Peter exported Claude's entire code base into ChatGPT, turned down Astra's initial suggestion to build the game in 3D using Blender, and instead had the model generate dedicated 2D illustrated art assets to transform *No Moat* into a visually rich, polished indie title.

[18:24](https://www.youtube.com/watch?v=iDrEXFOvFUc&t=1104s) Generating 2D Game Art and Stereotypical Tech Heroes

To kick off the visual design of *No Moat*, Peter instructed Astra to focus on building "one convincing fight first" with great graphics, specifically requesting the diorama aesthetic seen in his earlier gameplay demo. He highlights that one of the biggest advantages of working within ChatGPT and the broader OpenAI ecosystem is the native ChatGPT image model, which excels at generating game assets. The model produced artwork for the hero characters, the bad guys, the game stage, bugs, and items. Peter initially went off on a sidetrack to make the characters appear animated by generating 3D models in Blender, but because the 3D outputs didn't look anywhere near as detailed as the 2D illustrations, he decided to stick strictly to 2D.

When Peter prompted the AI to expand the roster of heroes, the results leaned heavily into tech industry tropes. He laughs at what he calls "probably the most stereotypical image of a PM and a designer I've ever seen"—pointing out the Product Manager sporting New Balance shoes while working with sticky notes, alongside an equally cliché designer, wryly noting that "AI can be pretty stereotypical." From there, he began playtesting and feeding notes back into the model, asking it to check gameplay balance, study proven games in the roguelike deck-builder genre for best practices, and generate more unique enemies and characters. Throughout this entire iterative loop, Peter emphasizes that he "did not look at the code a single time."

- **Hero archetypes generated:** Product Manager (wearing New Balance shoes and holding sticky notes), Designer, Vibe Coder, Principal Engineer, and a visibly burned-out Ex-FAANG Employee.

[20:17](https://www.youtube.com/watch?v=iDrEXFOvFUc&t=1217s) Shipping to ChatGPT Sites, Realistic Enemies, and Mobile Polish

Because *No Moat* is a 2D game rather than a 3D Godot project, Peter was able to deploy it directly to ChatGPT Sites. He points out an important configuration detail for anyone using the platform: newly generated web pages are restricted by default to people you explicitly share them with, so creators must click the "Share" button and change the permissions to **Public** so anyone on the web can play. Following deployment, he asked Astra for another polish and balance pass to make the game ship-ready, while also rethinking the enemy design. Initially, the AI had cast rival founders as the main antagonists, but Peter felt that wasn't true to life since "AI founders generally try to help each other." To make the opposition more realistic, he had the AI replace rival founders with industry headaches and generate custom portraits for each—including a "Predatory VC" that amuses Peter because the illustration looks just like the Joker.

For the final stretch of development, Peter wanted the game to run smoothly on both desktop web and mobile browsers. When he first tested the layout on his phone, it looked "pretty bad," so he showed the issue to Astra and prompted it to optimize the mobile UI and make additional improvements along the way. Altogether, turning *No Moat* into an actually playable game rather than a rough tech demo took about two hours of back-and-forth feedback. Because he genuinely loves the roguelike card-builder genre, he plans to keep iterating on it even just for his own enjoyment.

- **Updated enemy roster:** Scope Creeper, Procurement Manager, Reply Guy, and Predatory VC (styled like the Joker). - **Total development time:** ~2 hours of conversational iteration to reach a fully playable web and mobile game.

[21:59](https://www.youtube.com/watch?v=iDrEXFOvFUc&t=1319s) The Simple 3-Step Workflow: Brainstorm, Generate Art, Give Feedback

Wrapping up the tutorials, Peter reflects on how drastically the AI game-building workflow has changed in just a few months. Previously, he relied on rigid tools like "plan mode" and elaborate, heavily engineered prompts to coax AI into building games, and still wasn't really succeeding. In contrast, the games showcased in this video were built simply by exploring an idea conversationally, generating a reference image, and giving ongoing feedback to Astra.

```mermaid flowchart LR accTitle: Peter's Iterative AI Game Building Process accDescr: A three-step workflow showing how Peter builds games with GPT 6 Astra using casual voice dictation and feedback without looking at code. subgraph Ideation ["1. Ideation & Visuals"] A["Explore idea with AI"] --> B["Generate reference art"] end subgraph Iteration ["2. Conversational Loop"] C["Playtest & dictate feedback"] --> D["Astra updates game & balance"] D --> C end subgraph Deployment ["3. Release"] E["Ship to Godot / ChatGPT Sites"] end Ideation --> Iteration Iteration --> Deployment ```

Remarkably, half of the prompts he used were "just lazy voice dictation while I was out with my family." He explains that the underlying models have crossed a threshold of capability where "you can just talk to it like you talk to a capable game designer or developer, and they'll just get it done."

[22:34](https://www.youtube.com/watch?v=iDrEXFOvFUc&t=1354s) Honest Take on GPT 6 Astra's Launch Hype and Backlash

Peter concludes by addressing the controversy surrounding GPT 6 Astra's rollout. The launch sparked community backlash because OpenAI published a blog post, sent out press releases, and coordinated with influencers who boasted, "Hey, I've had access to this model for a week and here's how cool it is," at a time when no normal user could actually access it. In OpenAI's defense, Peter acknowledges that the delay may not have been entirely their fault, noting that third parties are likely involved in deciding when a frontier model is cleared for public release.

Even so, speaking from his perspective as a creator, Peter argues that "hyping this stuff up is completely unnecessary" and that launches should only happen the moment a model is genuinely available to everyone. With new models dropping almost every week, not every release is going to be life-changing, so he refuses to overhype them—preferring instead to offer his honest opinion and share practical builds that viewers can replicate themselves. Now that Astra is finally open to everyone, he encourages viewers to build the games of their childhood and imagination, adding that building games alongside your kids makes for "an amazing family activity."

## Notable quotes > [00:00](https://www.youtube.com/watch?v=iDrEXFOvFUc&t=0s) Astra is unequivocally the best model he has ever tested for game development — Peter Yang

> [01:58](https://www.youtube.com/watch?v=iDrEXFOvFUc&t=118s) Star Fox meets a collapsing megastructure — Peter Yang

> [06:00](https://www.youtube.com/watch?v=iDrEXFOvFUc&t=360s) Titanfall meets moving train — Peter Yang

> [11:46](https://www.youtube.com/watch?v=iDrEXFOvFUc&t=706s) better lighting effects — Peter Yang

> [11:46](https://www.youtube.com/watch?v=iDrEXFOvFUc&t=706s) StarCraft 101 — Peter Yang

> [18:24](https://www.youtube.com/watch?v=iDrEXFOvFUc&t=1104s) did not look at the code a single time — Peter Yang

> [18:24](https://www.youtube.com/watch?v=iDrEXFOvFUc&t=1104s) probably the most stereotypical image of a PM and a designer I've ever seen — Peter Yang

> [18:24](https://www.youtube.com/watch?v=iDrEXFOvFUc&t=1104s) AI can be pretty stereotypical — Peter Yang

> [20:17](https://www.youtube.com/watch?v=iDrEXFOvFUc&t=1217s) AI founders generally try to help each other — Peter Yang

> [21:59](https://www.youtube.com/watch?v=iDrEXFOvFUc&t=1319s) just lazy voice dictation while I was out with my family — Peter Yang

> [21:59](https://www.youtube.com/watch?v=iDrEXFOvFUc&t=1319s) you can just talk to it like you talk to a capable game designer or developer, and they'll just get it done — Peter Yang

> [22:34](https://www.youtube.com/watch?v=iDrEXFOvFUc&t=1354s) hyping this stuff up is completely unnecessary — Peter Yang

> [22:34](https://www.youtube.com/watch?v=iDrEXFOvFUc&t=1354s) Hey, I've had access to this model for a week and here's how cool it is — Peter Yang

> [22:34](https://www.youtube.com/watch?v=iDrEXFOvFUc&t=1354s) an amazing family activity — Peter Yang

## Mentioned GPT-6 Astra, Peter Yang, Star Fox, StarCraft, Dustline, Ashfall, No Moat, Blender, Godot, Blender MCP, Godot MCP, ChatGPT, ChatGPT Plus, ChatGPT Desktop, ChatGPT Sites, Model Context Protocol, OpenAI, Anthropic, Claude, Fable, Three.js, Blizzard Entertainment, Slay the Spire, Wildfrost, Left 4 Dead, Last Train Home, Titanfall

## Video description Copy my best AI workflows to save time and automate busywork: https://www.behindthecraft.com 🤝 Subscribe to my practical AI newsletter read by 150K+ professionals: https://creatoreconomy.so

GPT 6 Astra is the best model I’ve ever used for building games. In this tutorial, I walk through how to use Astra, Blender, and Godot to build a Star Fox-style space shooter, a moving-train first-person shooter, a StarCraft-style real-time strategy level, and No Moat, a roguelike deck builder about AI founders.

What I cover: (00:00) Astra makes my childhood gaming dreams come true (01:00) How to set up Astra, Blender, and Godot (02:00) Build a Star Fox-style space shooter (06:00) Build a first-person shooter on a moving train (09:30) Build a real-time strategy StarCraft level (13:30) Build a roguelike deck builder about AI founders (22:00) My simple process: Brainstorm, generate art, and give feedback (22:30) My honest take on Astra’s launch hype and backlash

Play No Moat here: https://no-moat.petergyang.chatgpt.site/

Where to find me: X: https://x.com/petergyang LinkedIn: https://www.linkedin.com/in/petergyang/

Subscribe to this channel - more tutorials coming soon!

[Source](https://www.youtube.com/watch?v=iDrEXFOvFUc)

中文简报

OpenAI 的 GPT-6 Astra 自动化操控 Blender 和 Godot 构建 3D 游戏通过 MCP 将 GPT-6 Astra 连接到 Blender 和 Godot，用简单的概念图和随意的语音反馈取代手动编程和复杂的提示词模板。 Peter Yang 展示了他如何在一个周五晚上，仅用每月 20 美元的标准 ChatGPT Plus 订阅服务，并在“中等”（Medium）设置下运行 GPT-6 Astra，就制作出了四款截然不同的游戏。通过模型上下文协议（MCP）将 ChatGPT 桌面应用连接到开源工具 Blender 和 Godot，AI 能够自主创建 3D 资产、编写游戏逻辑，甚至对三个 3D 演示项目进行试玩测试：一款《星际火狐》（Star Fox）风格的太空射击游戏、一款名为《Dustline》的移动列车第一人称射击游戏，以及一个受《星际争霸》（StarCraft）启发的即时战略（RTS）关卡《Ashfall》。在第四个项目中，他制作并部署了一款名为《No Moat》的 2D 肉鸽（roguelike）卡牌构筑游戏，以此讽刺 AI 初创公司生态；他将原本基于表情符号的原型从 Claude 迁移到了 ChatGPT，以利用其内置的图像生成功能和 ChatGPT Sites 托管服务。最后，Yang 强调，复杂的提示词工程已不再必要——创作者只需构思概念、生成参考图，并通过日常对话式的反馈（包括用手机进行语音口述）来引导 AI。 GPT-6 Astra 可通过模型上下文协议（MCP）连接器自主协同操控 Blender 和 Godot，使非技术背景的创作者无需编写代码或学习复杂的软件界面，即可构建可玩的 3D 游戏。 (1:04) 仅使用每月 20 美元的标准 ChatGPT Plus 订阅层级，就能在一个晚上构建出涵盖街机太空射击、第一人称射击、即时战略和卡牌构筑四个不同类别的游戏原型。 (0:00) 利用先进模型进行游戏开发的方式，已从死板的提示词模板转变为对话式语音口述、迭代反馈循环与集成图像生成的结合。 (21:59) 作者批评了 OpenAI 针对 Astra 采取的网红优先分阶段推出策略，认为 AI 产品只有在模型能同时向所有付费订阅者开放时才应正式发布。 (22:34)

Reading summary

Peter Yang argues that OpenAI's GPT-6 Astra is the most capable model yet for game development, demonstrating that non-technical creators can build playable 2D and 3D prototypes in under an hour by orchestrating external engines through conversational feedback. Orchestrating Blender and Godot via MCP (1:04) Rather than writing game code or navigating complex software manually, the workflow links the ChatGPT desktop application running GPT-6 Astra in medium reasoning mode directly to external creative tools. Yang utilized the Model Context Protocol integrations—specifically Blender MCP and Godot MCP—which allowed Astra to autonomously generate 3D assets in Blender, script game logic, and assemble playable scenes inside the Godot engine. Yang emphasizes that this pipeline requires no prior expertise in 3D modeling or game design software, nor does it necessitate higher-tier enterprise tiers like Astra High or Ultra. The standard $20-per-month ChatGPT Plus subscription provided sufficient reasoning capability to execute tool installations, direct asset generation, and compile multi-asset game scenes without manual coding. Prototyping 3D Shooters and RTS Mechanics (1:58) To evaluate Astra's mechanical versatility, Yang produced three distinct 3D prototypes in roughly thirty to sixty minutes each. These included a Star Fox-style rail shooter featuring multi-phase boss damage with destructible components, Dustline—a first-person shooter aboard a moving train with weapon swapping and jetpack adversaries—and Ashfall, a StarCraft-style real-time strategy slice featuring Terran marines, auto-healing medics, mobile siege tanks, and automated enemy swarm pathfinding. Achieving these results relied on iterative steering rather than one-shot generation. Yang utilized ChatGPT-generated concept art as visual anchors and prompted Astra for qualitative adjustments, such as asking for improved lighting effects to instantly elevate flat 3D environments. Furthermore, Astra leveraged browser automation and computer use to self-playtest builds, autonomously diagnosing gameplay mechanics and tuning late-wave difficulty curves across roughly a dozen prompt exchanges. 2D Asset Pipelines and Web Deployment (13:30) For No Moat, a roguelike deck builder satirizing Silicon Valley startups, Yang demonstrated an alternative 2D workflow. The initial logic and product specification were drafted in Claude using placeholder emojis before the codebase was migrated to Astra once public access opened. Rather than forcing low-fidelity 3D models, Astra generated custom 2D illustrations for tech archetypes—including Product Managers and Predatory VCs—and drafted card mechanics covering startup risks such as wrapper competitors and scope creepers. The entire deck builder reached a playable state in approximately two hours without Yang reviewing the source code once. Because the project was compiled for the web, it was published directly through ChatGPT Sites after switching permissions to public, with Astra resolving mobile browser layout issues following targeted user feedback. Conversational Feedback Versus Launch Hype (21:59) Yang highlights that the development paradigm has transitioned from rigid prompt engineering and structured planning modes to fluid, conversational iteration. Roughly half of his guidance for these projects was dictated via mobile voice mode while away from his desk, treating the model as an autonomous game designer capable of translating casual critiques into balance changes and gameplay tweaks. Addressing the launch controversy surrounding GPT-6 Astra, Yang critiques OpenAI's marketing rollout, noting that coordinated influencer campaigns before widespread public availability generated justifiable user backlash. While acknowledging that third-party release clearances may explain delays, he argues that companies should avoid unnecessary advance hype, urging creators instead to judge frontier models on replicable, practical utility. While Astra demonstrates remarkable speed in assembling playable vertical slices without manual programming, these projects remain early-stage prototypes rather than polished, commercial titles. Yang maintains that conversational toolchain orchestration fundamentally lowers the barrier to game creation, provided creators steer models with clear artistic references and realistic expectations.

Source quotes

[0:00] Peter Yang: Astra is unequivocally the best model he has ever tested for game development [1:58] Peter Yang: Star Fox meets a collapsing megastructure [6:00] Peter Yang: Titanfall meets moving train [11:46] Peter Yang: better lighting effects [11:46] Peter Yang: StarCraft 101 [18:24] Peter Yang: did not look at the code a single time [18:24] Peter Yang: probably the most stereotypical image of a PM and a designer I've ever seen [18:24] Peter Yang: AI can be pretty stereotypical [20:17] Peter Yang: AI founders generally try to help each other [21:59] Peter Yang: just lazy voice dictation while I was out with my family [21:59] Peter Yang: you can just talk to it like you talk to a capable game designer or developer, and they'll just get it done [22:34] Peter Yang: hyping this stuff up is completely unnecessary [22:34] Peter Yang: Hey, I've had access to this model for a week and here's how cool it is [22:34] Peter Yang: an amazing family activity

Original transcript

[0:00] Hey everyone, GPT Astra just launched. [0:03] And I could talk about how insane the [0:05] benchmarks are or hype it up like all [0:07] the other influencers, but instead, I [0:09] want to share how it's brought back a [0:11] part of my childhood. Because this is by [0:13] far the best model I've ever used for [0:16] building games. [0:18] In this video, I'm going to show you how [0:19] I created a Starfox style space shooter, [0:23] a first-person shooter on a moving [0:25] train, [0:26] a real-time strategy game like [0:28] Starcraft, and my favorite category, a [0:31] roguelike deck builder featuring an AI [0:34] founder trying to build a generational [0:36] company. [0:38] I worked on all four games at the same [0:40] time on Friday night using Astra on [0:43] medium. I think you can get quite far [0:45] even on the $20 ChatGPT Plus plan. You [0:47] really don't need Astra higher or ultra. [0:50] Medium can go a long way. [0:52] And at the end of this video, I'll also [0:54] talk about why the Astra launch got some [0:57] backlash and why I think it might not [0:59] have been entirely Open AI's fault. But [1:02] first, let me show you how to build some [1:03] games. [1:04] Okay, so here's the setup. To build [1:05] great games, you need the ChatGPT [1:07] desktop app. You need Astra selected [1:09] here [1:10] in the model selector, and you also need [1:13] two additional tools. [1:16] Blender is a free open-source tool to [1:19] create 3D models and animations. [1:22] And Godot is a free open-source game [1:24] engine that turns those models into [1:27] something that you can play. [1:29] And you know the funny thing is, I [1:30] actually have no idea how to use Blender [1:33] or Godot. I mean, just look at Blender's [1:35] UI. It looks very complicated. [1:37] And all I did was I just asked ChatGPT [1:40] to install Blender and Godot for me so [1:43] that AI can actually drive the apps [1:45] itself, right? Simple as just saying [1:47] let's install Godot MCP and Blender MCP. [1:51] All right, so now let me show you the [1:52] four games I built and how I built them, [1:55] starting with the Starfox style space [1:57] shooter. [1:58] So, Starfox is a space shooter from the [2:01] Super Nintendo days. And I really loved [2:04] playing it as a child. I want to start [2:06] with this because I've been using it to [2:08] test models ever since GPT 5.5. [2:12] Here's GPT 5.5's version of Starfox. As [2:15] you can see, it looks pretty great. It's [2:16] got the polygon graphics. It's got [2:18] enemies approaching me. And it looks [2:20] pretty fun. [2:21] Now, here is GPT 5.6's version. [2:25] And I've also compared it with Fable 5. [2:27] So, the graphics look a little bit [2:28] better. Look a little bit more polished. [2:30] There's also reactions here from your [2:32] wingmates and so on and so forth. And to [2:35] be fair, in these early demos, I just [2:37] used 3.js, which is a JavaScript library [2:40] for 3D in the browser. You can get quite [2:43] far with 3.js. But for Astra, I decided [2:45] to use Blender and Godot, too. And that [2:48] gives it a lot more tools to work with. [2:50] I can make the graphics a lot better. [2:53] Here is Astra's version. As you can see, [2:56] it looks a lot more impressive right off [2:57] the bat. Let's go ahead and play it. [3:00] So, I got the wingman. I got ChatGPT [3:02] images to generate some wingman profiles [3:04] for me. There's ship coming. The levels [3:06] are more interesting. I think there's [3:07] more ships here that I can attack. [3:10] And if you keep going, there's also [3:12] obstacles. There's different stages, [3:14] right? There's a stage over here. And [3:16] there's falling blocks that you have to [3:18] avoid. [3:19] And I also added a boss at the very end. [3:22] So, let's go ahead and get rid of the [3:23] ships. [3:24] And here it is. This is kind of like a [3:26] starship destroyer kind of boss. And if [3:28] you shoot at it, [3:30] different parts of the ship will [3:32] hopefully break apart. So, let's try [3:33] this. [3:35] All right. So, let's shoot at it. So, [3:36] here you go. We broke the wings. [3:39] And the other wings. And you keep [3:40] shooting at it. And eventually, the boss [3:43] is going to die. All right. So, that is [3:46] basically the game that I built. And [3:47] now, let me show you how I built this, [3:49] right? [3:51] So, I asked ChatGPT and Astra to install [3:53] Godot and Blender, and it installed. And [3:56] then I pointed it to this blog post that [3:58] an OpenAI employee wrote [4:01] on how to build games with Astra. So, [4:03] basically, Astra read the blog post, and [4:06] the best practices are [4:08] describe the player experience first, [4:09] establish a look with concept images, [4:12] build repeatable play tests early, and [4:14] use Blender for distinctive assets. [4:17] Okay. So, then I asked it, "What kind of [4:18] visually impressive and super fun games [4:20] should we make? Give me your top five [4:22] ideas." And it gave me a bunch of ideas. [4:25] And then I decided to just tell it to [4:27] build Star Fox. Okay. So, then it gave [4:29] me some more ideas. Star Fox meets a [4:31] collapsing mega structure, which is what [4:33] we ended up building, the game that I [4:34] just ended up showing you. [4:36] So, then I just told it to go ahead and [4:37] start building it, right? [4:39] Then it built some sort of game that [4:41] actually was not that fun. And I [4:43] actually got ChatGPT in another thread [4:46] to generate images for what the game [4:49] could look like. And this is the image [4:50] of what the game could look like. All [4:52] right. So, you can see this is a [4:53] collapsing space structure. There's [4:54] obstacles along the way. [4:56] And I asked it to add teammates and [4:59] allies and make the graphics look [5:01] better. Then it kind of went off and did [5:03] a lot more work. It generated a [5:05] representative example of what the what [5:07] the game should look like here. [5:09] And it also generated a profile images [5:11] for my wingman. [5:13] And then I just started giving Astra me [5:14] more feedback, right? So, the game looks [5:16] better, but it's still boring. Add [5:18] power-ups, make the enemies harder, the [5:20] space station make it look like it's [5:22] falling apart. [5:23] And then the boss fight should be more [5:25] interesting as well. [5:27] And another thing that I did is like I [5:28] just asked it, "Hey, how would you [5:30] continue to improve this?" And it gave [5:32] me a bunch of ideas here. [5:34] And then I asked it to make the boss [5:36] more intimidating and implement all of [5:38] its ideas. [5:40] One thing that Astra is really good at [5:41] is it can use browser use and use [5:43] computer use to play test its own game [5:46] and just make sure that it's fun. Uh so [5:49] there you go. So, after basically 3 [5:51] minutes of conversation, we ended up [5:53] with the game that I just showed you. [5:55] Now, let me show you the next game, [5:57] which is a first-person shooter on a [6:00] moving train. [6:02] So, this game is called Dustline. And [6:04] let's go ahead and play it. [6:06] All right. So, you see here that I'm on [6:08] the cart of a moving train. There's [6:09] enemies shooting at me. There's gun. I [6:11] can reload. And also, I can aim the gun. [6:15] Like this. [6:16] And looks like there's a flamethrower. [6:17] Let's go and pick up the flamethrower. I [6:19] can shoot flames at the enemies. [6:22] Also, rocket launcher. Let's shoot the [6:24] rockets at these guys. [6:26] All right. It makes bubble explosions. [6:29] Let me switch back to the gun. That [6:30] gun's the best, I think. [6:31] Let's get rid of these guys. [6:33] And always with these games or maybe [6:36] tech demos I like to build, I always [6:38] like to add a boss battle at the end. [6:40] So, let's get rid of these guys first. [6:42] And let's see what the boss looks like [6:43] at the very end of this game. [6:45] All right. [6:47] Here we go. Some of the enemies Okay, [6:49] some of these enemies fly. They have [6:51] jetpacks. They just shoot at me. [6:53] All right. Hopefully, I don't die before [6:54] I get to the boss. [6:55] And you kind of see the boss already, [6:57] right? It's basically like a huge ship. [6:58] All right. And I think for this boss, we [7:00] probably should get some rocket [7:01] launchers. [7:03] Flamethrower is probably not great. [7:05] Let's try using a rocket launcher to [7:07] kill it. [7:08] And let's see what happens after we kill [7:09] it. [7:10] All right. The boss is not really [7:11] shooting at me. Maybe I need to make it [7:12] harder. [7:13] Looks like I ran out of rocket ammo. [7:16] There we go. [7:17] And let's go here, take control of the [7:19] train, and that's the game. Okay? I [7:21] mean, look. It's really more of a tech [7:22] demo than a real game. But once again, [7:24] let me show you how I built this. [7:26] So, I followed the same process, right? [7:28] The first process is, "Hey, let's build [7:30] a first-person shooter with Godot and [7:32] Blender. Give me your top five ideas." [7:34] And it gave me these five ideas. [7:37] And honestly, I thought about maybe [7:38] doing some sort of Left 4 Dead zombie [7:40] shooter. [7:41] Kind of explore that. [7:43] And then one of the ideas was the last [7:45] train home idea. And that kind of [7:47] inspired me to maybe make this a moving [7:49] train shooter instead. [7:52] So then I got ChatGPT generate an image [7:54] of what it's going to look like. [7:55] Now, the image looks a lot more [7:57] impressive than the game that we built, [7:58] even though the game that we built is, [7:59] in my opinion, pretty good. So maybe the [8:01] next version of OpenAI's model can [8:03] actually build this in one shot. But [8:05] it's always good to provide a reference [8:06] image so that AI knows what you're going [8:08] for, right? [8:10] And then here's the idea, Titanfall [8:12] meets moving train. [8:14] It loved the idea. [8:16] And I asked it for a three-milestone [8:18] plan. So you can try to get it to build [8:19] in one shot, but I think it's good to [8:21] kind of split it into like a couple of [8:22] milestones so they can test it along the [8:24] way. [8:26] And it built something fairly basic, [8:28] okay? It didn't quite look like the way [8:30] I wanted it to. So I provided a [8:32] reference image here again. And I told [8:34] it to make the graphics look better, [8:36] make the gun look more detailed, and [8:38] instead of robots as the enemies, make [8:40] it actual humans. [8:42] And add proper death animations and [8:43] stuff. [8:44] Okay? [8:45] Then it kind of went off and built it [8:47] again, and I provided more feedback, [8:49] right? And this time I think I actually [8:50] had to leave. I had to go out with my [8:52] family. So I started providing feedback [8:54] using my voice and ChatGPT on the mobile [8:56] phone. [8:57] So this is like super stream of [8:58] conscious feedback. Have more different [9:01] types of enemies. The AI is pretty dumb [9:03] right now. This is not like super [9:04] quality feedback, right? And I also told [9:06] it to make the rocket launcher better [9:08] and add more bullets to it. [9:10] And so on and so forth. And that's it. [9:12] That ended up with the game that I just [9:14] demoed to you guys, right? [9:16] Okay, so the reason I want to go through [9:18] these conversations is I really want to [9:21] emphasize that anyone can do this, as [9:23] long as you have ChatGPT, Astra, and you [9:26] have Blender and Godot MCP installed, [9:28] which are free and open source. [9:31] The three-step process is just [9:32] brainstorm some ideas, get ChatGPT [9:34] images to to some concept art, and then [9:37] ask AI to build and give it feedback [9:38] along the way. And this whole process [9:40] took me maybe like 30 minutes to hour to [9:42] do. [9:43] While at the same time building three [9:44] other games. [9:46] All right, for the next game I want to [9:47] show you [9:48] is basically related to my favorite game [9:50] of all time, which is StarCraft. You [9:53] know, I played StarCraft non-stop in [9:54] high school. I skipped classes to play [9:56] it. [9:57] And funny story, my defining moment in [9:59] StarCraft was playing with my Korean [10:01] friend three versus one stomping a [10:03] computer, and then I dropped a nuclear [10:05] bomb on his base and kind of ruined his [10:08] perfect record. That That's my favorite [10:10] story. So, there's rumors that Blizzard [10:13] is working on a new StarCraft game [10:15] coming out this fall. [10:16] So, naturally I had to see if Astra [10:18] could help me build a StarCraft level. [10:21] All right, so let me show you the game [10:22] that it built. So, the game is called [10:25] Ashfall, and let's join the fight here. [10:28] And you can see here that this is a [10:29] real-time strategy game. I got marines [10:31] that I can control, and I got siege [10:32] tanks. [10:33] And the siege tank can actually siege if [10:35] you ever played StarCraft before, and we [10:36] can build marines, medics, and tanks [10:38] here. And let's also build some turrets [10:40] while we're at it. [10:42] So, it's basically like a base defense [10:44] game. There's a zerg coming at me, and I [10:46] got to capture the relays [10:48] and defend myself against the zerg. [10:51] So, next wave is coming soon. We got [10:52] some medics that can heal my marines. [10:55] And you can see the controls work pretty [10:56] well, right? And I can also build some [10:57] turrets, and here's another tank. Let's [10:59] put it into siege mode. [11:01] And finally, I think there's other [11:03] options here like orbital, which I think [11:05] it's like a orbital strike. [11:07] Okay, that's pretty cool. And I can also [11:10] add stem to my marines. [11:12] I can press E to put them into [11:14] overdrive. Let's wait for the next wave [11:15] and see what happens. [11:17] All right, here we go. I press E to put [11:18] into overdrive, and then they'll kill [11:20] the enemies faster. Oh, and here's the [11:22] ultralisks that came out. So, I probably [11:24] should go back here. Okay, it looks like [11:25] I'm actually losing right now. There's [11:27] >> [laughter] [11:27] >> too many zerg. I think my base is [11:28] falling apart a little bit. Maybe I can [11:30] use the orbital strike again. There we [11:32] go. So, I survived this level, but I [11:33] probably won't survive the next level. [11:35] All right, so that's the game. I I want [11:37] to see if Astra can kind of build a [11:39] real-time strategy game, and this is [11:41] again just a cool tech demo, but I think [11:43] it's pretty good, right? So, now let me [11:44] show you how I built this. [11:46] So, [11:47] in this example, I knew that I wanted to [11:49] build StarCraft. So, I asked it to [11:50] create an image of a real-time strategy [11:52] game like StarCraft, except with modern [11:54] graphics, and this is what it made. And [11:57] once again, the concept art [11:59] looks a lot better than actual game that [12:00] we built, but I think it's useful to [12:02] ground the AI on what we're looking for. [12:05] Okay, so after it made the concept art, [12:07] I basically just asked it to make a [12:08] real-time strategy level with Blender [12:10] and Godot. Player control the human [12:12] Terran faction, and the enemy is the [12:14] Zerg, and has to be super fun, you know. [12:17] Once again, I feel like as the models [12:19] get better and better, I become more and [12:20] more lazy with my prompts. [12:22] So, it went off and started building the [12:23] stuff, [12:25] and then I said, "Hey, make the graphics [12:27] better, and why don't we do a level [12:28] where it's about Terran defending the [12:30] base against Zerg." And then I mentioned [12:33] that the siege tank obviously has to be [12:34] able to enter siege mode. This is like [12:36] StarCraft 101. [12:38] Okay, then it made the game, and making [12:40] a fun game isn't just about the [12:41] graphics, obviously. It's about balance, [12:43] it's about core loops, it's about [12:44] progression systems. So, I told it, "You [12:47] made the game too easy. Add some more [12:49] additional characters like medic, [12:51] and then add three different types of [12:53] Zerg enemies, [12:54] and make the enemies more aggressive." [12:56] Okay, and then I asked it for better [12:57] lighting effects. So, quick tip, I'm not [13:00] a game designer, but I found that if you [13:02] want to make your 3D games look good, [13:03] just asking for better lighting effects [13:06] makes a huge difference. [13:07] If you scroll down here, and then I play [13:10] tested the game myself, and I told it [13:11] that wave five onwards is way too hard. [13:13] Make it easier, and so on and so forth. [13:17] And that's it. So, I maybe took like [13:19] a dozen or so prompts to do this, and [13:22] it's obviously not StarCraft, but it's a [13:24] pretty a tech demo for what a real-time [13:27] strategy game made with AI could look [13:29] like. [13:30] All right. So, the last game I want to [13:31] show you guys is in my opinion the [13:33] closest to an actual game. [13:36] It is a rogue-like deck builder called [13:39] No Moat. [13:41] And basically, it's about the AI founder [13:42] who's trying to build a generational [13:44] company and trying to overcome obstacles [13:46] along the way. So, let's go ahead and [13:48] try playing it now. [13:50] So, we got three choices. We got a [13:51] dropout, the professor, and the ex-Fan [13:53] employee. And why don't we pick the [13:54] dropout? [13:56] And the first thing you see is a map of [13:58] the levels that you have to beat. [14:00] And [snorts] there's three different [14:01] stages. [14:02] And let's go ahead and get a companion [14:04] first. So, we can get a designer, a [14:06] principal engineer, or a research [14:07] scientist. [14:09] I'm going to have to go with a designer [14:10] to start. [14:11] And here is our first enemy fight, [14:13] right? So, basically, it's a rogue-like [14:15] deck builder. I got a bunch of decks [14:16] here, and I can go ahead and fight the [14:18] enemy. [14:20] Let's go ahead and just use Ship It to [14:21] get rid of some these rapper startups, [14:23] and let's get our designer on the board. [14:26] All right. Let me rate limit this scope [14:28] creeper [14:29] so that it doesn't actually attack. And [14:31] Ship It, get rid of this thing. [14:34] And yeah, there you go. So, there's a [14:35] new player coming, and I can also press [14:37] this to refresh my cards. [14:39] So, let me just go ahead and play it for [14:40] a little bit longer. [14:42] Let's add a GPU rack here to block the [14:44] enemy's attacks. [14:46] And Ship It again. [14:48] And there you go. Here's the mini boss, [14:50] the Acquire, that wants to acquire our [14:52] startup before we even got started. So, [14:54] let's go ahead and try to beat it. Let's [14:56] move our players around. [14:57] And let's hit this thing, and let's do [15:00] hype tweet to increase our attack. All [15:02] right. [15:03] And since it's got three health left, [15:05] let's just kill it directly with Ship [15:06] It. [15:07] There you go. So, we won the game, a [15:09] victory victory. [15:10] And now we have a choice between adding [15:12] more companions or getting more cards. [15:13] So, let's say adding more companions. [15:16] And this time we want to add [15:18] a research scientist. We're going to [15:19] have a research scientist. [15:21] All right. And here's a bug [15:22] investigation. [15:24] And again, we can kill the bugs, [15:27] bring our pieces onto the board, [15:29] and see what other enemies are coming [15:31] out. So, we got a memory leak. [15:33] Let's use Hypetweet to kill it. [15:36] And yeah, so that's basically the game. [15:39] There's different levels, and every [15:41] level has a little mini boss boss fight. [15:43] So, now there's like a may major bug [15:45] that we have to beat. [15:46] And unlike the previous tech demos that [15:48] I showed, I just spent a lot more time [15:50] into this game to make it good. Because [15:52] as you know with roguelike deck [15:54] builders, they got to look great, but [15:56] they also have to be somewhat balanced, [15:57] right? [15:58] And by the way, for this game, I'm going [16:00] to include a link for you to play it [16:02] directly in the YouTube description and [16:03] the first comment. I'm actually pretty [16:05] proud of it, and please give me [16:06] feedback, and I'm going to improve it [16:08] over time. [16:10] All right, so for this game, I actually [16:11] started in Claude using Fable, but [16:14] that's only because I didn't have access [16:16] to Astra for the first day. [16:18] So, you can pretty much do this using [16:20] Astra and ChatGPT. But basically, I [16:22] followed the same process. First thing [16:24] is I have to build a game like Slay the [16:26] Spire and Wildfrost. It needs to be [16:28] super fun. [16:30] Ideally, there's some memes about the AI [16:32] race in there, too. And give me your [16:34] five best ideas. And the very first idea [16:36] that they came up with is No Moat. [16:38] You run a AI lab climbing a ramp from [16:40] seed round to AGI. [16:42] Units are researchers and GPU clusters, [16:45] and there's different types of cards. [16:46] So, then it recommended a bunch of other [16:48] ideas, [16:49] and I told it that I want to do No Moat. [16:52] So, then I asked it like, is this very [16:53] complicated to build? And it said it's [16:55] pretty easy, and here's a rough plan. [16:57] And this time I actually told it to [16:58] write an HTML spec for the game. [17:01] And let me kind of show you what the [17:02] spec looks like. [17:04] Okay, so here's the spec that AI [17:06] generated. I have a spec scale where I [17:08] split the spec into product design and [17:10] tech stack. Right, so product here is [17:12] the solution, here's the core loop, [17:14] and here's the requirements, [17:16] and design. [17:17] This is what it looks like. So, the [17:19] thing about Claude, the big kind of [17:21] drawback, is that Claude doesn't have an [17:23] image model. So, in terms of making the [17:25] designs, it has to generate stuff [17:28] probably using HTML and using emojis. [17:31] So, the first version of the game [17:32] actually just looks like this with [17:34] emojis and different cards, just to kind [17:36] of get the gameplay and get a feel of [17:38] the game. [17:39] Right? So, here's the design, here's the [17:41] emojis, and so on. [17:43] And then I kind of realized that I [17:44] really wanted better images than just [17:46] emojis. So, after playing the first [17:48] version, I decided to move over to [17:50] ChatGPT and Astra became available then, [17:52] too. [17:54] So, let's go ahead and move over to [17:55] ChatGPT [17:56] and let me show you the thread. [17:58] In ChatGPT, Claude had already built the [18:00] initial version of the game with emojis. [18:03] So, I told it to first export the code [18:04] base and build a playable version of the [18:07] game. [18:08] And [18:10] I said, "Hey, should we use Blender MC [18:12] for this? I have no idea what I'm doing, [18:13] so I need your help." [18:15] And it actually didn't recommend using [18:17] Blender. [18:18] It it recommended using images to just [18:20] generate art. [18:22] And scrolling down here, [18:24] and I say, "Hey, why don't we figure out [18:26] one convincing fight first? [18:28] Let's make the graphics great. And I [18:30] really like the diorama look that you [18:32] just saw in the demo that I just [18:33] played." So, I kind of gave it [18:35] instructions to do that. [18:36] And so on and so forth, and here we go. [18:39] And you can see here that I started [18:40] generating the artwork. So, this is [18:42] like, again, one of the biggest benefits [18:44] of ChatGPT and the whole OpenAI [18:46] ecosystem, the ChatGPT image model, it's [18:49] just great for generating game art. [18:51] So, it generated both the bad guys [18:54] and my hero characters. [18:56] And it also generated the game stage, [18:58] some of the bugs, and I items, and so [19:00] on. [19:02] Okay? And then I really wanted each [19:03] character to appear animated. So, we [19:05] kind of went off on a side track. We [19:08] tried to use Blender to generate the [19:09] characters, but the 3D characters didn't [19:11] look anywhere near as detailed as this [19:14] 2D art. So, I decided to stay on the 2D [19:16] side. [19:18] All right. Then, I asked it to generate [19:19] more heroes. So, you have some [19:22] characters here and then you have [19:24] probably the most stereotypical image of [19:26] a PM and a designer I've ever seen. So, [19:28] this is the PM. It's got the new balance [19:30] shoes. It's doing sticky notes. [19:32] And the designer looks something like [19:33] this. [19:35] Right? Uh so, yeah. So, AI can be pretty [19:37] stereotypical. [19:38] Then, I started playing the game and [19:40] just give me some more feedback. Just [19:41] like asking it to check the balance, [19:43] look at other proven games in this [19:45] genre, and see if there's any best [19:47] practices from them. [19:49] And also generating more characters, [19:52] more enemies, asking it to make each [19:54] enemy unique, and so on and so forth. [19:56] And obviously, I did not look at the [19:57] code a single time in this whole [19:59] process. [20:00] So, I asked it to generate a vibe coder [20:02] and a principal engineer. [20:04] And going down here. [20:06] And I told it to add the X fan employee [20:09] as one of the hero characters. [20:12] All right. And then here's the [20:13] illustration for the X fan employee. [20:15] Looks pretty burned out. [20:17] And finally, I told it to ship this game [20:19] in ChatGPT sites. So, unlike the 3D [20:22] games, which are in Godot, like this is [20:24] a 2D game, so it's pretty easy to ship [20:26] to ChatGPT sites. And then I said you [20:28] can play it here. So, real quick with [20:30] ChatGPT sites. [20:31] Once you generate the web page, [20:34] by default, it's only available to [20:36] people you share with. So, you got to [20:38] hit share here and change the [20:39] permissions to public so that anyone on [20:42] the web can play it. [20:44] Then, I asked it to fix some more stuff [20:45] and so on and so forth and then do a [20:47] polish and balance pass [20:49] and make it ship ready. [20:50] And there you go. Oh, another thing that [20:53] it did was [20:54] a lot of the enemies were rival [20:56] founders, but I think AI founders [20:58] generally try to help each other. So, I [21:00] wanted to make the enemies more [21:01] realistic. So, now the enemies became [21:03] stuff like [21:05] Skull Creeper, Procurement Manager, [21:07] Reply Guy, so on and so forth. [21:09] And of course, I also got AI to generate [21:12] images for each enemy type. [21:15] I think this is the predatory VC. It's [21:17] pretty funny. Looks like the Joker. [21:19] Yeah. Oh, the last thing I told it to do [21:21] is I wanted to make it playable on both [21:23] web and mobile. So, I tested it on [21:26] mobile and it looks pretty bad, as you [21:27] can see here. So, I asked it to make it [21:29] look better on mobile [21:31] and just make a bunch more improvements [21:33] along the way. [21:35] And there you go. That is the game. [21:37] So, just like the other examples, it's [21:39] just like a lot of back and forth, [21:41] giving it a feedback, and making it [21:43] better over time. For this one, it took [21:45] me maybe like 2 hours of back and forth [21:47] to make a actually playable game instead [21:49] of a tech demo like this. And like I [21:52] said, I love this genre of roguelike [21:54] card builders. So, I'm going to keep [21:55] improving this game and make it better [21:57] even just for myself. [21:59] All right. So, to wrap up, it's funny [22:01] because just a few months ago, I was [22:03] using things like plan mode and [22:05] elaborate prompts to try to get AI to [22:07] build games and not really succeeding. [22:10] But as you can see here, most of these [22:11] games I built from just exploring an [22:13] idea with AI, generating a reference [22:16] image, and giving feedback to Astra [22:18] along the way. [22:20] And half of the prompts that I showed [22:21] you were just lazy voice dictation while [22:24] I was out with my family. [22:26] The thing is, these models have become [22:27] so good that you can just talk to it [22:29] like you talk to a capable game designer [22:31] or developer, and they'll just get it [22:33] done. All right. So, now I want to talk [22:34] about Astra's launch. [22:37] So, the launch actually got a little bit [22:39] of a backlash because what happened was [22:42] OpenAI made a blog post and did press [22:44] releases and did influencers saying, [22:46] "Hey, I've had access to this model for [22:48] a week and here's how cool it is." And [22:49] the problem was no normal person could [22:52] get it. And in OpenAI's defense, this [22:54] may not have been entirely their fault. [22:56] As you know, there are probably third [22:58] parties involved in deciding when a [23:00] state of the model can get released. [23:03] But, personally as a creator, I think [23:05] hyping this stuff up is completely [23:07] unnecessary. I think launches should [23:09] happen only when the model is actually [23:11] available to everybody. [23:14] The reality is that there's so many [23:15] models coming out almost every week now [23:17] that not every model is going to change [23:19] your life. So, I don't want to overhype [23:21] this stuff. Instead, what I want to do [23:23] is just give you my honest opinion [23:26] and share what I build with these models [23:29] so that you can try building them, too. [23:32] So, now that we all have access to [23:33] Astra, I encourage you to build the [23:35] games from your childhood and also your [23:37] imagination. [23:38] Build them with your kids, too. It's an [23:40] amazing family activity. [23:42] And as always, thanks for watching and [23:45] please like and subscribe if you enjoyed [23:46] this video.