Ghostyx

loot engineer

[Ghostyx @0xGhostyx]

@0xKense1 i am not sure, but looks like /compact?

[Ghostyx @0xGhostyx]

Long game builds are the boss fight for Claude Code limits. One chat, 30+ turns, and every call drags the whole history along.

Kensei measured it on 945 runs. After a 30-turn session, /clear cut the next questions by 79%, with the same right answers.

Patch notes for my next build:
- new feature, new session: /clear
- same bug, long hunt: /compact
- game design rules in a skill, not in CLAUDE.md

Do you respawn your sessions, or play the whole run in one chat?

[Ghostyx @0xGhostyx]

Long game builds are the boss fight for Claude Code limits. One chat, 30+ turns, and every call drags the whole history along.

Kensei measured it on 945 runs. After a 30-turn session, /clear cut the next questions by 79%, with the same right answers.

Patch notes for my next build:
- new feature, new session: /clear
- same bug, long hunt: /compact
- game design rules in a skill, not in CLAUDE.md

Do you respawn your sessions, or play the whole run in one chat?

[Kensei @0xKense1]

https://t.co/WUk7rWCYuZ

[Ghostyx @0xGhostyx]

@0xKense1 so quick tips, but i am so lazy to do even them 😂

[Ghostyx @0xGhostyx]

@0xKense1 Wow, really huge difference

[Ghostyx @0xGhostyx]

lauren tan, engineering lead of grok bot, told peter yang on behind the craft when you can let a bot work without watching it:

"it ultimately comes back to trust, right? like how much trust are you willing to give your agents or your bots"

her loop: watch the bot do the job, correct it, turn it into a skill, then a routine.

· 00:33 - it comes back to trust
· 00:58 - turn the session into a skill
· 01:46 - from a one-shot skill to a routine
· 02:28 - why it has to be gradual
· 02:55 - the full flow for a new grok bot user

i gave grok bot 5 real tasks with hidden tests: 10 of 10. full breakdown in the article below

[Ghostyx @0xGhostyx]

lauren tan, engineering lead of grok bot, told peter yang on behind the craft when you can let a bot work without watching it:

"it ultimately comes back to trust, right? like how much trust are you willing to give your agents or your bots"

her loop: watch the bot do the job, correct it, turn it into a skill, then a routine.

· 00:33 - it comes back to trust
· 00:58 - turn the session into a skill
· 01:46 - from a one-shot skill to a routine
· 02:28 - why it has to be gradual
· 02:55 - the full flow for a new grok bot user

i gave grok bot 5 real tasks with hidden tests: 10 of 10. full breakdown in the article below

[Ghostyx @0xGhostyx]

https://t.co/LdFa1NrE8V

[Ghostyx @0xGhostyx]

@0xKense1 yeap, now it is clear

[Ghostyx @0xGhostyx]

https://t.co/LdFa1NrE8V

[Ghostyx @0xGhostyx]

@0xRewind thx bro

[Ghostyx @0xGhostyx]

https://t.co/LdFa1NrE8V

[Ghostyx @0xGhostyx]

@0xKense1 so bullish, great job!

[Ghostyx @0xGhostyx]

John Carmack, who co-created Doom and Quake, was CTO of Oculus and now runs the AGI research company Keen Technologies, explains why AI that beats humans at Atari still can't learn a game the way you do.

"When people undertake working in these modern games, the temptation is so great to look beyond the pixels and look at internal data structures. And honestly, I still just look at that as cheating."

· 00:00 - 200 million frames: a month of nonstop play
· 00:56 - Give the same agents two hours and they stop looking impressive
· 02:06 - Why not Pokemon or Minecraft
· 02:32 - Reading game data instead of pixels is cheating
· 03:09 - Reality is not a turn-based game

My Doom mashups were the easy side of this. Claude wrote mods that talk to the game's internals directly. Playing Doom from pixels alone, the way he wants agents to learn, is a much harder problem.

[Ghostyx @0xGhostyx]

John Carmack, who co-created Doom and Quake, was CTO of Oculus and now runs the AGI research company Keen Technologies, explains why AI that beats humans at Atari still can't learn a game the way you do.

"When people undertake working in these modern games, the temptation is so great to look beyond the pixels and look at internal data structures. And honestly, I still just look at that as cheating."

· 00:00 - 200 million frames: a month of nonstop play
· 00:56 - Give the same agents two hours and they stop looking impressive
· 02:06 - Why not Pokemon or Minecraft
· 02:32 - Reading game data instead of pixels is cheating
· 03:09 - Reality is not a turn-based game

My Doom mashups were the easy side of this. Claude wrote mods that talk to the game's internals directly. Playing Doom from pixels alone, the way he wants agents to learn, is a much harder problem.

[Ghostyx @0xGhostyx]

https://t.co/B01YlSawrd

[Ghostyx @0xGhostyx]

@0xRewind thx bro

[Ghostyx @0xGhostyx]

https://t.co/B01YlSawrd

[Ghostyx @0xGhostyx]

https://t.co/B01YlSawrd

[Ghostyx @0xGhostyx]

@0xKense1 nice setup

[Ghostyx @0xGhostyx]

Epic Games, the studio behind Fortnite and Unreal Engine, shows in 8 minutes how Claude Code builds a game level inside Unreal from text prompts.

"What we've shown here today would take months to build by hand. But with the MCP Server and Unreal, our artists were able to make all of this in days."

· 00:00 - Why making games is slow
· 00:52 - Claude Code plugged into Unreal
· 01:38 - A living room from prompts
· 03:00 - Scaling up to a whole city
· 04:34 - Why models don't think in 3D on their own
· 05:39 - Where the model gets it wrong
· 06:51 - A hazard zone in days, not months

I gave two Claude models the same one-line ski game prompt. Haiku 5.5 made a playable but empty slope. Opus 5.5 made gates, jumps and a score.

Epic says it on stage too: "models will still do some really unpredictable things." Pick your model like you pick your loadout.

[Ghostyx @0xGhostyx]

@0xKense1 glad you liked it

[Ghostyx @0xGhostyx]

Epic Games, the studio behind Fortnite and Unreal Engine, shows in 8 minutes how Claude Code builds a game level inside Unreal from text prompts.

"What we've shown here today would take months to build by hand. But with the MCP Server and Unreal, our artists were able to make all of this in days."

· 00:00 - Why making games is slow
· 00:52 - Claude Code plugged into Unreal
· 01:38 - A living room from prompts
· 03:00 - Scaling up to a whole city
· 04:34 - Why models don't think in 3D on their own
· 05:39 - Where the model gets it wrong
· 06:51 - A hazard zone in days, not months

I gave two Claude models the same one-line ski game prompt. Haiku 5.5 made a playable but empty slope. Opus 5.5 made gates, jumps and a score.

Epic says it on stage too: "models will still do some really unpredictable things." Pick your model like you pick your loadout.

[Ghostyx @0xGhostyx]

same idea
two Claude models
one prompt each

Haiku 5.5
· a skier and a slope
· trees and rocks
· playable, but empty

Opus 5.5
· mountains, gates, jumps
· speed, distance, score
· feels like a real game