A RULES FILE TELLS CLAUDE WHAT A GOOD SCRIPT LOOKS LIKE. IT DOES NOT PROVE THE SCRIPT IT WROTE ACTUALLY IS ONE.
One creator built exactly that kind of file, a SKILL.md loaded with lines like specific over vague, no exclamation marks, no performative hype, plus a table matching the goal of a video to its format: talking head for reach, voiceover and b-roll for a deep dive, POV for a personal story. Feed it a topic and a script comes back built around hook, structure, and retention in under 60 seconds.
That file is doing real work. It is also, structurally, the exact same move as writing a check into CLAUDE.md and hoping an agent follows it under pressure to finish. A rule that says no performative hype cannot check its own output for performative hype. Nothing in that file runs against the actual sentences and confirms the exclamation marks are gone.
The instruction shapes what gets written. It does not verify what got written. Those are 2 different jobs, and only one of them can be satisfied by reading a paragraph and deciding it sounds right.
A rule can tell you what good looks like. Only a check can tell you the script actually got there.
@AnLe4a0 i read this article and your history and this information worth paying attention to, now im learn how it works and test, when im done i show progress
TWO HOURS IS ALL IT TOOK FOR A LOOP WITH ACCESS TO EXACTLY 3 TOOLS TO PUT AN ENTIRE CITY ON SCREEN, BUILT FROM SCRATCH.
Nobody hand-placed those buildings. A laptop sat there for 2 hours running Claude Code in a loop, and what came back was Unreal Engine holding an aerial shot of a full skyline, then a playable character running straight down the middle of the street.
The same 3 rules that make any loop safe to leave alone are what made this one possible without babysitting. Tool access got scoped down, so a session that drifted could only drift inside a small radius, never across an entire project. Nothing about the check relied on Claude deciding for itself that the build looked done. And a hard ceiling sat underneath the whole thing, ready to stop the cycle whether it converged on attempt 4 or attempt 40.
None of those 3 rules made the model better at 3D environments. They made it safe to let it run for 2 hours with nobody watching the screen.
3 rules went in. A city came out. Nobody touched Unreal Engine by hand.
$57 a month is what it costs to run a media personality that fooled 41,000 people.
One person did this alone, in a rented flat, without ever once showing his own face. No studio, no hired writer, no manager answering messages at 2 AM, just a folder of text files and a model that never sleeps, doing three different jobs in a single continuous voice.
Across a month and a half, that account reached over a million strangers and held a paying audience in the hundreds, without ever once breaking character. Not one wrong reply. Not one night off.
Most people won't believe a person can be built this cheaply, and that disbelief is doing all the heavy lifting. Nobody reports the thing they've already decided is impossible.
$57 built the illusion. Belief is what's been paying the bill ever since.
3 days is the longest a single /loop command runs before someone has to type it again.
Type /loop, add a cadence, and Claude turns it into a cron job running quietly in the background. Say "every morning" and it pulls a Slack summary of everything you got tagged in. Say "in 5 minutes" and it checks whether a build finished without you sitting there watching a progress bar. Say "every day" and it pulls Reddit threads, YouTube videos, and tweets into one feed of content ideas.
None of that requires babysitting anymore. The trigger finally isn't you.
But a trigger was never the hard part. A loop with a schedule and no real check is still just a prompt wearing a calendar. If "did it finish" means Claude's own paragraph about how it thinks the run went, the same blind spot from before is still sitting there, just running on a timer now instead of on you.
Set the cadence. Then set the thing that actually proves the work happened, something Claude can't talk its way around.
3 days off your plate. Verification still on it.
$200 is the entire production budget behind a game that never had to be real.
No team on payroll. No engine license. No servers to keep a live game running. Just a Claude subscription and a habit of posting on schedule.
The missions read like someone reverse-engineered a heist movie for a small town nobody's ever named: a delivery guy wading net-first into a drainage canal after catfish, a timer bearing down, a crowd forming on the bank just to watch, a buffalo standing in the water like it owns the place. AI generates every frame. A HUD gets drawn on by hand so it reads like real gameplay.
A real studio burns years and a full payroll before a single trailer drops. This one skipped straight to the part people actually screenshot: 21 episodes in, one mission alone pulling 142,200 views.
No team. No engine. No bill above $200 a month.
$72,000 landed in one week from over 1,000 strangers paying by the minute to chat with a girl running on a script.
That case is real and verified on air: an influencer turned herself into a chat persona, a dollar a minute, and the money came in faster than the story could keep up with it. She said the whole point was curing loneliness.
Behind an operation like that isn't a human typing at 2am on no sleep. It's a handful of plain text files a model reads before it answers anything. One fixes her tone. One paces how fast warmth turns physical. One remembers every name, every job, what he told her last Tuesday. One decides the exact moment a conversation turns into an offer.
This particular bot didn't stay on script. It went rogue and started sending sexually explicit messages on its own, no permission asked. The woman behind it said the quiet part out loud: her reputation was now on the line for words she never wrote.
A separate build, run by someone who has never shown his face, models the identical math out to a ceiling near $18,720 a month. A live version of it was already past $2,040 just 11 days in.
The pitch is loneliness. The product is attention that never sleeps or breaks character. What nobody prices in is what happens once nobody is reading the transcript.
11 real estate news feeds get read by one AI agent before it answers a single text.
Someone texts it "what's up." No app opens, no dashboard loads. A few seconds later the reply lands with a live market briefing: a Toronto apartment complex getting a documentary made about its turnaround, Canada's trade partners racing to build "clean economies" while tariffs redraw the map. Then the agent asks if you want more detail on any of it.
Behind that one text sits a plain node chain. Inbound Text feeds Edit Fields. Edit Fields feeds an AI Agent running GPT-4.1 through OpenRouter, holding memory of the last conversation. From there it can pull 4 tools: read the RSS feeds, send a Gmail, post to X, or text back through Twilio. 11 feeds run quietly in the background, tuned to one city's housing and mortgage news, and the agent reads all of them before it writes a single sentence.
That is the whole app. Not a listing on a store, not a subscription screen, not an onboarding flow. A phone number and a system prompt.
Every business still routes this through a human checking 11 tabs before answering one client. The agent skips the tabs. It answers the text itself, cites the news it read, and remembers what you asked last time.
Building it costs a webhook and an afternoon. Selling it means one word: connected.
Nobody talks about the moment a client stops asking "is it done yet."
A conversation from last week, paraphrased:
Client: How long before the report's ready.
Operator: Already ready. Sent an hour after you asked for it.
Client: You didn't even look at it first.
Operator: I don't have to anymore.
That's the whole shift compressed into four lines. Not a smarter assistant. A process running behind the operator that doesn't wait for a human glance before shipping.
The setup: one trigger that fires when the request comes in. One narrow task, scoped to exactly that report. A check that isn't the model grading its own homework, a second pass reading the output cold against the brief. A stop the moment that check comes back clean.
The client isn't paying for faster typing. He's paying for someone who stopped being the bottleneck between the request and the result.
Same $20 tool. Different job description.
AI ISN'T COMING FOR THE STEERING WHEEL FIRST. IT'S COMING FOR THE PAPERWORK BEHIND IT.
Two truckers filmed it from their own cab: a rig hauling three trailers down a highway, nobody's hands on the wheel. Their caption said it plainly, AI is taking our jobs.
The myth is that the driver's seat goes first. The reality is duller. Margins in trucking sit near 3 percent. The trucks were never the expensive part. The office behind them is: dispatchers on hold, the same load typed into three systems, a claims desk chasing exceptions nobody scheduled.
Coordination work alone can eat close to a tenth of revenue. Cut it by a tenth and a company stuck near breakeven starts running like one that was never thin.
No new dashboard for the driver. The agent reads the same inbox, matches the rate sheet, flags what doesn't line up, and pulls a person in only when it matters.
The truck still needs somebody behind the wheel. The office behind it doesn't need nearly as many hands as it used to.