An NPC in GTA 6 got cut off in traffic.
She followed the player, got out of the van, walked up with what looks like a knife…
then backed off and drove away.
No scripted panic.
It plays like an agent:
trigger → target → push → stop.
Rockstar reportedly built ~600k animations.
The part people actually share is the decision.
This is the same test we use for AI agents outside games.
We stop calling it a bot the second it picks a goal and knows when to quit.
Defuse or Provoke?
🤖 Made with AI
2
3
187
A "GTA VI cheat codes" list is going viral today. It's fake.
Claude checked it. The first three lines are GTA Vice City codes from 2002, word for word. The rest are scrambled old GTA cheats with typos like "Garbage Track".
What's actually real on launch day:
• GTA 6 has 61 cheats, but you have to earn every one
• Infinite ammo = 25 silver medals at shooting ranges
• $250,000 = finish the story first
• Only 17 codes are public so far
Full list of the 17 known codes in the images.
+ GTA 5 PC & phone codes.
Would you grind 100 workouts for body armor, or just play it straight?
A "GTA VI cheat codes" list is going viral today. It's fake.
Claude checked it. The first three lines are GTA Vice City codes from 2002, word for word. The rest are scrambled old GTA cheats with typos like "Garbage Track".
What's actually real on launch day:
• GTA 6 has 61 cheats, but you have to earn every one
• Infinite ammo = 25 silver medals at shooting ranges
• $250,000 = finish the story first
• Only 17 codes are public so far
Full list of the 17 known codes in the images.
+ GTA 5 PC & phone codes.
Would you grind 100 workouts for body armor, or just play it straight?
Code, build it yourself: github.com/f0rgent/iphone-ma…
Star it if you build one. Which Apple feature should Claude Code build next?
More: f0rgent.github.io
56
HOLY SH*T
Claude Code stopped working one tool call at a time. With Code Mode, Claude writes JavaScript that runs the whole chain in one step. I went through the docs, and the gotchas matter more than the launch post. The short version:
→ turn it on: CLAUDE_CODE_EXPERIMENTAL_CODE_MODE=1 claude (needs v2.1.323+)
→ one script runs commands in parallel, passes each result to the next call and hands Claude only the part it needs
→ use it for edits repeated across many files, searches that need filtering, and commands that can run at once
→ Ctrl+O shows the code Claude wrote behind the tool calls
→ in auto mode it can run in the background for up to 2 hours while you keep talking. otherwise it dies at 10 minutes, and time you spend on permission prompts counts
→ your deny rules and PreToolUse hooks still check every call inside the script
→ but the auto mode classifier reviews the code, not each call, and can't see paths computed at run time. keep hard rules for anything you can't afford
→ want to audit the script itself? a PreToolUse hook on REPL gets the full JavaScript in tool_input.code
→ a Code Mode session only resumes with the variable still set. forget it and Claude Code refuses to resume
→ it won't start with DISABLE_TELEMETRY set, and Remote Control, Chrome and computer use are off while it's on
→ first test: "Read every Markdown file in this project and list the links that point to local files that don't exist."
The job shifts from approving every tool call to reviewing the script that makes them.
Save this before you flip it on.
Check out Code Mode, a new, experimental way for Claude Code to work with tools. Claude writes scripts that chain your tools together in sophisticated ways, and it can answer you while long commands are still running.
Opt in with CLAUDE_CODE_EXPERIMENTAL_CODE_MODE=1 code.claude.com/docs/en/code…
HOLY SH*T
Claude Code stopped working one tool call at a time. With Code Mode, Claude writes JavaScript that runs the whole chain in one step. I went through the docs, and the gotchas matter more than the launch post. The short version:
→ turn it on: CLAUDE_CODE_EXPERIMENTAL_CODE_MODE=1 claude (needs v2.1.323+)
→ one script runs commands in parallel, passes each result to the next call and hands Claude only the part it needs
→ use it for edits repeated across many files, searches that need filtering, and commands that can run at once
→ Ctrl+O shows the code Claude wrote behind the tool calls
→ in auto mode it can run in the background for up to 2 hours while you keep talking. otherwise it dies at 10 minutes, and time you spend on permission prompts counts
→ your deny rules and PreToolUse hooks still check every call inside the script
→ but the auto mode classifier reviews the code, not each call, and can't see paths computed at run time. keep hard rules for anything you can't afford
→ want to audit the script itself? a PreToolUse hook on REPL gets the full JavaScript in tool_input.code
→ a Code Mode session only resumes with the variable still set. forget it and Claude Code refuses to resume
→ it won't start with DISABLE_TELEMETRY set, and Remote Control, Chrome and computer use are off while it's on
→ first test: "Read every Markdown file in this project and list the links that point to local files that don't exist."
The job shifts from approving every tool call to reviewing the script that makes them.
Save this before you flip it on.
Check out Code Mode, a new, experimental way for Claude Code to work with tools. Claude writes scripts that chain your tools together in sophisticated ways, and it can answer you while long commands are still running.
Opt in with CLAUDE_CODE_EXPERIMENTAL_CODE_MODE=1 code.claude.com/docs/en/code…
Keep it on for every session, ~/.claude/settings.json:
{ "env": { "CLAUDE_CODE_EXPERIMENTAL_CODE_MODE": "1" } }
More tested Claude Code setups: f0rgent.github.io
4
98
WHAT THE F*CK
Claude just rejected 25 of 31 frames of a video I'd already posted. I stole the Claude Code team's verification rules for it. The short version:
→ never let the model that made it grade it. a second Claude caught what the first one shipped
→ grade stills, not the vibe. 31 frames pulled from a 92s video, each scored 1 to 10
→ a frame scores as low as its worst flaw. one stacked caption made a 2/10
→ check every caption against its shot. mine said "DOOM ON iDRIVE" over a hand on a gear lever
→ check the boring stuff. Russian UI slipped into an English post
→ every note has to be a fix you can make: hard cut, zoom in, cut the shot
→ grade again after every fix. round 2: worst frame went from 2 to 5
→ when an edit can't fix it, stop and call a human. my hand covers the screen, so: reshoot
→ ship at 8/10 or don't ship
The critic prompt I used is in the first reply. Steal it.
WHAT THE F*CK
The Claude Code team now has Claude write and review ~90% of its code. This post is how they keep that from breaking. The short version:
→ verification loops are what let you run dozens of Claudes in parallel instead of 1-2
→ a separate grader checking against a rubric raised task success by up to 10 points
→ if you pick one test type, pick end-to-end. a screenshot of the broken step tells Claude more than “Timeout after 30s”
→ property tests throw thousands of inputs at rules that can’t break, like “balance never goes negative”
→ differential tests run the old and new system on the same inputs and flag any difference
→ a /verify skill in .claude/skills/ boots and drives the real app before every commit
→ still failing after 3 rounds, Claude stops and calls a human
→ a change can never edit the check that judges it
→ every commit is stamped with how it was verified, even “none”
The job shifts from reading every diff to defining what “done” looks like.
11
11
405
the critic prompt. paste it into a new chat, never the one that made the thing.
want the full loop with contact sheets and red boxes? frame-critic is free, with 3k+ other Claude skills: f0rgent.github.io
1
87
I can't f*cking believe Fable 5.2 just did this.
Gave it one job: find the drone and take it down. No human aimed a single shot.
It spotted the drones, locked on, swung a 3D-printed turret, put a laser on the target, and fired.
4 drones. One dropped. One took a hit and kept flying.
Built on a desk at home. No lab. No defense budget.
Stop using Fable as an expensive chatbot. Give it eyes and a trigger.
Would you let an AI pull the trigger on its own?
24
27
717
the tracker part as one prompt 👇 paste it into Claude Code with Fable 5.2, plug in any camera.
more builds like this: f0rgent.github.io
86
holy shit. my son's crayon drawing just turned into a 3,500-drone show.
this is the most dad thing I've ever built with AI.
he drew a goofy elephant and a peacock. I built a prompt that takes any kid's drawing and turns it into a full drone light show plan, then fed his picture to Fable 5.2.
it found both animals
put a drone on every crayon line
kept his colours
checked every flight path every 0.1s, 0 collisions
wrote a flight file for all 3,500 drones
the sky at the end is what a show like this looks like for real.
from a crayon on paper to this. I'm still not over it.
what would your kid draw?
Full prompt in the first reply. Bookmark it before yours asks for a dragon.
holy shit. my son's crayon drawing just turned into a 3,500-drone show.
this is the most dad thing I've ever built with AI.
he drew a goofy elephant and a peacock. I built a prompt that takes any kid's drawing and turns it into a full drone light show plan, then fed his picture to Fable 5.2.
it found both animals
put a drone on every crayon line
kept his colours
checked every flight path every 0.1s, 0 collisions
wrote a flight file for all 3,500 drones
the sky at the end is what a show like this looks like for real.
from a crayon on paper to this. I'm still not over it.
what would your kid draw?
Full prompt in the first reply. Bookmark it before yours asks for a dragon.
here's the full prompt 👇 attach any drawing and paste it into Claude Code.
more prompts like this: f0rgent.github.io
1
133

















