
The new roadster is literally a rocket car. This pretty much writes itself Built this meme game on @spawn in a few hours, complete with ceiling driving... spawn.co/@sidriff/roadster-a…

Gameplay designer. 15+ years of taste. Iterating with AI to find the fun much faster.

The new roadster is literally a rocket car. This pretty much writes itself Built this meme game on @spawn in a few hours, complete with ceiling driving... spawn.co/@sidriff/roadster-a…


Fascinating research for any developer in this space. Lots of disclaimers and nuance in article (please, please, please don't start pumping cheap naval 4x games). My read is players who like more open-ended, emergent experiences care less about how the sausage gets made.
Are Steam players of particular genres less accepting of AI? newsletter.gamediscover.co/p…

I read Ethan’s book Co-intelligence in 2024 and it had a profound impact on the way I saw and interacted with LLMs. Game development seemed like a perfect application. Small projects need implementation muscle. Big projects need help with maintaining coherence. Frontier intelligence has demonstrated wonderful capabilities for solving both problems. But only if you have your eyes open to both the problems and the possibilities. Highly recommend following Ethan if you want to broaden your perspective.
For those who don’t follow video games, there is constant policing of any AI use among small, indie developers. They are the most resource constrained firms in a field where profits are rare & artistic vision is often compromised, but they are punished more harshly than big devs.

Approximately 100 prompts later... I took some feedback and...am I accidentally remaking Homeworld? Would that even be so bad...?
I decided to test-drive Grok 4.6 by building this little space fleet auto-battler in three.js (directly in the Grok Build beta). The premise was to build a battle worthy of cinema, with enough moving parts to make direct control impossible. Read on for details or experience it yourself: spacefleet-3d-proto.grok.me This took 260 prompts. I prefer to steer the agent bit by bit as we work through features rather than turn over a giant spec for autonomous building. It's more time consuming, but as a creative there's nothing more exhilarating than the freedom to iterate with the speed of Grok 4.5/4.6 and watch your dreams come to life in front of your eyes. Roughly, it went down like this: 1 - Concept. I started by having Grok generate some concepts using some references from games and movies. 2 - Scene Layout. Once we found a direction that I liked, we sketched out the battle scene in 3D and hammered out some painful early mistakes (lighting, hitching, backwards ships). Arranging the battle lines and getting the general feel of the standoff was the priority. 3 - Camera Grammar. I knew this wouldn't be very interesting unless we had a cinematic camera director. I tried out a number of modes, and kept adding them as the project matured. Includes wide/establishing shots, up close shots of weapons and damage, hull cams, missile follow cams, focus shots on breaking ships falling into the planet's gravity well, wing/formation shots, fly-bys, and a super-weapon showcase. 4 - Title Card and flow into the battle. We added some simple music (Grok is surprisingly good at this?!). The Star Wars-esque pan down to the planet. Warp-in easing and timing. 5 - Fleet differentiation. I didn't want the battle to be a mirror match. Could have taken this farther, but each side has its own configurations of weapons and roles in combat. The Vesper uses more beam weaponry and fighter screens. Kharis prefers to fight with missiles and laser barrages. 6 - Juicy Death. Capitals don't just pop, they sink toward the planet and bristle with explosions while ejecting gas and armor plates. The camera director prefers to show case this when possible. I also put some love into missile interception; fighters can pick these off opportunistically. Pretty cool to see a missile get popped right before impact. 7 - Fighters ate the most turns by far, and they're my favorite part of this experiment. I wanted more than a busy hornets nest of particles, so these units intelligently cluster while they engage with enemy fighters, intercept missiles, and strafe capital ships. They have a proper RCS-style maneuvering paradigm complete with rolling-pitch turns and evasive barrel roles. 8 - Escape Sequence. Rather than fight to death, the capitals will try to disengage and warp away from battle (sometimes dying in the process). If the flagship goes down or bugs out, everyone else scrambles (complete with fighter coverage). 9 - Debug and Tuning. Technically didn't happen last. Most of the diagnosis happened by looking at stats on the battle engine and camera director. There were a few moments in there where I needed to look into hard numbers or output, so there's some visualization left in there (O and L) primarily to debug what's going on with fighters. If you got this far, thanks for reading. I've got a lot of ideas on where to go next with this, but I'm happy to hear feedback and suggestions.

Some notes on what didn't go so well, possibly some feedback for SpaceXAI. 1 - Context window management. As far as I can tell, there isn't any in the app builder. When the project gets more complicated, the agent tends to truncate or wipe files. You can nudge it to split large files which tends to help. 2 - Saving progress is generally cumbersome. Pushing to Github is clunky and sometimes it fails. That makes issue #1 extra terrifying. So far I've not encountered a state I couldn't recover from. 3 - All of these app builders always spring for React/R3F and in my experience it always gets in the way. There's some obvious stutter with the faster moving pieces, even on my 240hz display. I'm guessing once I rebuild it in straight three.js, that will go away. 4 - Sound. It's always the hardest part. I have some workflow solutions that involve existing libraries/elevenlabs but I was trying to stick with app builder and I wasn't willing to paste my API key in the chat! Disappointed that Build mode doesn't have access to the Grok Voice API (but it can get to Grok Imagine). All in all, I'll probably stick with Grok Build in the terminal. But build mode makes it so easy to get up and running with an idea.

I decided to test-drive Grok 4.6 by building this little space fleet auto-battler in three.js (directly in the Grok Build beta). The premise was to build a battle worthy of cinema, with enough moving parts to make direct control impossible. Read on for details or experience it yourself: spacefleet-3d-proto.grok.me This took 260 prompts. I prefer to steer the agent bit by bit as we work through features rather than turn over a giant spec for autonomous building. It's more time consuming, but as a creative there's nothing more exhilarating than the freedom to iterate with the speed of Grok 4.5/4.6 and watch your dreams come to life in front of your eyes. Roughly, it went down like this: 1 - Concept. I started by having Grok generate some concepts using some references from games and movies. 2 - Scene Layout. Once we found a direction that I liked, we sketched out the battle scene in 3D and hammered out some painful early mistakes (lighting, hitching, backwards ships). Arranging the battle lines and getting the general feel of the standoff was the priority. 3 - Camera Grammar. I knew this wouldn't be very interesting unless we had a cinematic camera director. I tried out a number of modes, and kept adding them as the project matured. Includes wide/establishing shots, up close shots of weapons and damage, hull cams, missile follow cams, focus shots on breaking ships falling into the planet's gravity well, wing/formation shots, fly-bys, and a super-weapon showcase. 4 - Title Card and flow into the battle. We added some simple music (Grok is surprisingly good at this?!). The Star Wars-esque pan down to the planet. Warp-in easing and timing. 5 - Fleet differentiation. I didn't want the battle to be a mirror match. Could have taken this farther, but each side has its own configurations of weapons and roles in combat. The Vesper uses more beam weaponry and fighter screens. Kharis prefers to fight with missiles and laser barrages. 6 - Juicy Death. Capitals don't just pop, they sink toward the planet and bristle with explosions while ejecting gas and armor plates. The camera director prefers to show case this when possible. I also put some love into missile interception; fighters can pick these off opportunistically. Pretty cool to see a missile get popped right before impact. 7 - Fighters ate the most turns by far, and they're my favorite part of this experiment. I wanted more than a busy hornets nest of particles, so these units intelligently cluster while they engage with enemy fighters, intercept missiles, and strafe capital ships. They have a proper RCS-style maneuvering paradigm complete with rolling-pitch turns and evasive barrel roles. 8 - Escape Sequence. Rather than fight to death, the capitals will try to disengage and warp away from battle (sometimes dying in the process). If the flagship goes down or bugs out, everyone else scrambles (complete with fighter coverage). 9 - Debug and Tuning. Technically didn't happen last. Most of the diagnosis happened by looking at stats on the battle engine and camera director. There were a few moments in there where I needed to look into hard numbers or output, so there's some visualization left in there (O and L) primarily to debug what's going on with fighters. If you got this far, thanks for reading. I've got a lot of ideas on where to go next with this, but I'm happy to hear feedback and suggestions.



Anyone else building custom design tools or review UIs for agents? Tools are cheap enough to build now that I’ve settled into workflows that let me look at everything side-by-side. Generating unit and building concepts with Grok Imagine, then having the coding agent turn them into meshes. At each phase I left feedback affordances right in the tool. Keep/revise/reject plus a “copy JSON” button so I can just hand it over in terminal. Lately I’ve been playing around with side-by-side views so the iterating agent can hit both the concept and readability from the actual game camera. Feeling good about how this one’s coming together.

So much Opus 5 hype, lately. But what's actually INSANE is sleeping on Grok 4.5... I built this three.js mobile RTS in Grok's App Builder ON MY PHONE, no gauntlet loops or $5,000 token bills needed. Just simple design specs and iteration. Features Grok can build: - 3d assets + animation - sound and visual effects - music - videos (for the faction advisors) - UI - multiplayer - player bots - pathfinding - and much, much more

This is the way. Don't just stop after the /loop craps the game out. Iterate your way to a good experience.
Started with the character and core animations like the ollie on flat ground. Once that worked, ran ultracode on the map with AAA graphics overnight, with /loop and an agent verifier until done. After that, lots of iteration on game feel, tricks and physics…


Grok 4.5 remains my favorite iteration partner as a game designer. Speedy, accurate, always up for a good laugh. Grok Ops: grok-ops.pages.dev/ Claude of Duty: claude-of-duty.pages.dev/ Full changelog and source: github.com/sidriff/Grok-Ops Enjoy..!

Introducing Claude of Duty: Grok Ops Forked the impressive Opus 5 one-shot and spent an afternoon with Grok Build turning the scaffold into a proper playable firefight. New mode, allies, and menu. Then a LOT of little things. Is it a good game now? You tell me:
Claude Opus 5 one-shotted this game. EVERYTHING you see in this demo is custom code... not a single external asset was used. AI games are going to be amazing. (sound on)

For the curious, check out the prompt/demos/source here: camelot-comparison.pages.dev…
I'm cooking up a visual target in three.js for a low-poly castle siege game. Here's Grok 4.5, Opus 5, and GPT-5.6 Sol with identical prompts. I can share the full prompt and source code if anyone really wants it.