蜃気楼

AI, crypto @NeuroClubAi

[蜃気楼 @fuxps32]

He built a 5-level AI second brain. He runs his whole business on level 2.

The whole feed chases the top. Knowledge graphs. Vector databases. Always-on agents that sync while you sleep. He built every one of them to see what they did.

His real system is one Claude Code project called HERC2. Folders and markdown files, nothing else. At the root sits a claude.md that works as a router: who he is, how he works, which folder holds what. The agent reads it first and stops asking him where things live.

Level 1 finds a file by its name.
Level 2 pulls a whole topic into an LLM wiki.
Level 3 searches by meaning through a vector database.
Level 4 maps entities in a knowledge graph: Jordan works at Acme, Acme competes with a rival.
Level 5 is G-brain, the always-on memory Gary Tan built at Y Combinator, syncing on a schedule he never touches.

He tested all 5. Then he walked back down to 2.

Because when he asked a vector database to summarize his March 5 meeting, it grabbed 5 chunks and missed the other 15. A markdown file his agent reads top to bottom never misses. Boring is beautiful.

So he runs a Grillme skill that interviews him until it knows everything, drops the answers into markdown, and lets auto-memory in claude.md hold the rest.

Codex reads the same files through an agents.md copy.

Everyone screenshots the glowing graph. He ships from a folder of plain text.

[Moysei @0xMoysei]

https://t.co/Fnh6MxF3pS

[蜃気楼 @fuxps32]

The creator of Claude Code thinks software engineering jobs start disappearing by the end of this year. Then he described the loop he built to replace part of his own job, and why he still does it by hand.

Boris Cherny runs the fastest-growing AI coding tool in the world. On the podcast, he didn't soften the timeline. Engineering roles, he said, start going away soon.

But the human moment came at the end.

The most repetitive part of his day is answering users on X and Threads. Broken builds, feature gripes, tips. So he automated it. A loop in Claude Code, now a routine that runs every 30 minutes, pulling feedback through both APIs.

Then the line that stuck.

He has it fully automated. He prefers to do it himself.

The part he could hand to an agent is the part he won't. Because the complaints are where the product gets better. Nobody builds something great in one stroke of genius. It improves a little every day, and only if someone keeps listening.

He built the machine to skip the work. He kept the work on purpose.

[Moysei @0xMoysei]

https://t.co/4pjQifJHoT

[蜃気楼 @fuxps32]

A $9,000-a-month customer canceled a SaaS in 1 sentence.

The founder thought the product was finished. 40 features. 60 templates. A dashboard where you dragged blocks into the shape you needed. Everyone who saw the demo said it looked powerful.

The customer who left was the kind every founder wants. A power user, 7 months in. On the exit call he didn't complain about price or a missing feature.

He said he'd rather open Claude and build the exact thing himself in an afternoon.

That was the whole reason. The product made him do the arranging. Pick the blocks. Lay them out. Learn the system. In 2026 that's the part nobody wants to do anymore.

So the founder spent 3 weeks tearing it down. He killed the template gallery. He killed the block library. He gave the tool 1 input: you talk to it. You hold whisper, say "track this differently," and the screen reshapes around how you think, live, in seconds.

Not a tool with settings. Clay you mold by speaking.
The next power user never asked how it worked. He told it what he wanted and watched it become that.

The builders losing right now aren't shipping fewer features. They're shipping software you still have to operate.

In 2026 you don't hand people blocks. You hand them clay that listens.

[obssnnn @0xObssnnn]

https://t.co/AgqrgcP7gG

[蜃気楼 @fuxps32]

Anthropic's lead for Claude Code and Cowork:

You stopped being an IC. You became the manager of your agents.

In a 90-minute episode on Lenny's Podcast, Fiona Fung shows how she uses Claude to manage and review team output.

The loop changes shape. You delegate the task, the agents produce, you review the output like a manager reviews a team. Same skill, fewer keystrokes, faster cycles.

So the leverage moves. Not how much you can build, but how much output you can review and steer at once.

Manage the loops. Don't sit inside them.

[Moysei @0xMoysei]

https://t.co/4pjQifJHoT

[蜃気楼 @fuxps32]

Head of Claude Code at Anthropic:

"I can now run Claude for hours, or for days."

In a 40-minute fireside chat at Meta, Boris Cherny laid out the loop behind it.

It starts with dynamic workflows. Claude writes a small program inside a virtual machine, and that program orchestrates dozens, hundreds, thousands of other Claudes to finish the work.

The fuel is test-time compute, the 4th scaling law. 5 effort levels, low to max. More tokens out, better the result.

The lock came off with Auto mode. No yes/no on every command. Claude approves its own steps, prompt-injection success holds near 1% over 100 attempts, and the agent runs unattended for days.

The output: 1,700 PRs this year, 400,000 lines added, 8 billion tokens since March. 0 lines written by hand.

More signal than a $500 agents course. Save this.

He stopped coding. He started running the loop.

[Moysei @0xMoysei]

https://t.co/4pjQifJHoT

[蜃気楼 @fuxps32]

Most agencies charge thousands to get a business more Google reviews. He charges $5000 a month and lets one app do all the work.

The idea hit him in a parking lot. Half the local spots he passed had four star ratings and barely any reviews, the kind of thing that quietly kills walk in traffic.

Owners know it hurts them, but chasing reviews by hand is the last thing they have time for. So he sold them the fix instead of the chore.

He opened Emergent, typed a single prompt, and got a full app: a dashboard, automated texts, a direct review link, and a separate login for each business. He styled it like a Mercury dashboard so it looked worth paying for, then hooked it to Zapier and Square. Now it runs itself. A customer pays, two hours later a text lands with a one tap review link. The owner does nothing.

Every restaurant becomes a $500 line that renews on its own. The analytics page shows them the reviews climbing, so they never cancel.

Watch how he built the whole thing before someone in your area beats you to it.

[obssnnn @0xObssnnn]

https://t.co/AgqrgcP7gG

[蜃気楼 @fuxps32]

Loop engineering explainer: "Most people overcomplicate it."

Boris Cherny and Peter Steinberger both said the same thing: "stop prompting your agents, start writing loops." One creator went deep and broke a working loop into 4 blocks.

Trigger. What starts it. /loop runs on your machine, /schedule runs in the cloud, or wrap the whole thing in one orchestration skill you fire by name.

Execution skills. The saved, battle-tested skills that do the work. His rule: never build a loop without them, because they hold how you actually want each task done.

Goal and verification. A goal plus a rule that proves it is done. For fuzzy tasks, bridge the abstract to approved or not, or a 1 to 10 score, and let a separate agent grade it to cut bias.

Output and memory. Write results and lessons to a markdown file. As Addy Osmani put it, the agent forgets, the repo doesn't.

One guardrail before you set it loose: training mode. Make it pause for approval the first few runs so it does not burn tokens doing the wrong thing.

Save this. A loop without verification is just an expensive way to be wrong faster.

[Moysei @0xMoysei]

https://t.co/4pjQifJHoT

[蜃気楼 @fuxps32]

Creator of Bun: "This entire thing was one prompt, and it just ran for 30 minutes."

At Anthropic's Code with Claude, Jarred Sumner runs Claude in auto mode, no permission prompts, then walks away. On stage one prompt ran 30 minutes alone, found a bug, fixed it, and opened a PR on a Bun issue with 20 up votes.

The unlock is verification. Give the model a target, a way to test itself, and auto mode, and it climbs until it hits the goal. He says Opus 4.7 is the first model actually good at this.

He runs hundreds of these every night while he sleeps.

He stopped writing the code. He just writes the prompt that starts the loop.

[the lich @thelichhh]

https://t.co/v7WZdGPCaO

[蜃気楼 @fuxps32]

You don't need the most expensive model. You need a loop that forces a cheap one to perform like one.

One builder is proving it in the open. His system, millrace, turns any workflow into a governed loop, and it builds those loops using loops it already made.

Right now one of them is writing a compiler that pushes small models far above their size through harness engineering. Another is doing something stranger: fixing bugs in the loop framework itself. The thing improves the thing that improves it.

The terminal tells the story. A run moves through builder, checker, updater, arbiter, auditor, and planner, each stage gated, routed, and corrected on its own until it passes. Nobody types continue.

He has run this since January, and it is open source.

[the lich @thelichhh]

https://t.co/v7WZdGPCaO

[蜃気楼 @fuxps32]

The man who built Claude Code has a CLAUDE.md file with exactly 2 lines in it.

Most people stuff their config file with hundreds of rules. Boris Cherny keeps his at 2 lines. One auto-merges his pull requests once accepted.

The other posts them to a team channel so someone can stamp it and unblock him. Every other rule lives in a shared file the whole team edits a few times a week.

His advice when yours gets bloated: delete it and start fresh. People over-engineer this, he says, because the model changes with every release. Do the minimal thing that keeps it on track, then add back a line at a time only when it drifts. With each new model, you add less.

He also stopped trusting his own instincts. A teammate found a memory leak faster by asking Claude than Cherny did by hand with a heap dump and DevTools. Claude wrote its own tool to analyze the dump and found the leak first.

The skill is no longer knowing the rules. It is knowing how few you need.

[the lich @thelichhh]

https://t.co/v7WZdGPCaO

[蜃気楼 @fuxps32]

Two AI founders said stop writing prompts, write loops. One creator did the math and said most people cannot afford to.

Boris Cherny and Peter Steinberger both went public: they no longer prompt their coding agents, they write loops. The crowd ran with it.

A loop is three things. A trigger (a new pull request, or a timer every 15 minutes), an action, and a stop condition. It runs, saves to memory, and on the next trigger prompts itself for the next step. It repeats to the stop condition, or forever if you set it wrong.

The catch he names:

No project is fully defined on day one, so the loop builds a pile nobody asked for. The code turns ugly, since the agent leaves all 9 failed fixes in before the 10th works.

And on a $100 plan you cannot burn tokens the way founders with unlimited budgets do.

His verdict: the future, but not yet. Loops on top of vibe coding today is a 50-story tower built overnight on no foundation.

Loops are not new. The rich just got a name for them first.

[the lich @thelichhh]

https://t.co/v7WZdGPCaO

[蜃気楼 @fuxps32]

One engineer at Anthropic stopped working his own bug queue. It clears itself now.

He launched voice mode across the company's products, set up a routine, and walked away. It listens for every ticket, every GitHub issue, every bug report that mentions voice mode.

When one lands, it writes the fix and opens the pull request on its own. Boris Cherny, the man who built Claude Code, says he has never once talked to that engineer about how it works. It just runs.

The trick that lets a loop run that long is one rule. When Claude makes a mistake, the engineer does not correct it in the chat. He writes the correction into a CLAUDE.md or turns it into a skill. Patch it in the chat and it breaks again tomorrow. Write it into a file and it never repeats. Do that enough and the loop runs forever.

Cherny lives the same way. Whenever he needs code, Claude writes it. Whenever he needs a review, Claude runs it. Whenever he needs a security check, Claude does it. He talks to a loop, and the loop prompts Claude for him.

The engineer is still on the team. His feature has not needed him in months.

[the lich @thelichhh]

https://t.co/v7WZdGPCaO

[蜃気楼 @fuxps32]

He shipped 150 pull requests in a day. He typed 0 of them.

At a builders event the host asked the room who had Claude Code Psychosis. Boris Cherny, the man who built Claude Code, said his own team teases him that he has it.

He has not written a line of code by hand in 2026. Not one. The model writes 100% of it now.

The scale is the part that lands. On a normal day Cherny runs 5 to 10 sessions, and each one spins up its own swarm of agents. A few hundred run at once. On one of those days he shipped 150 pull requests, and his hands never touched the code.

The newer habit is loops. He runs dozens of them with /loop. One clusters his Twitter feedback every 30 minutes and reports back. His Claudes talk to each other over Slack while they run, sorting out the unknowns without him.

He reaches for the printing press to explain it. Books once got 100x cheaper, and literacy climbed from 10% to 70%. He thinks software is next, and faster.

He runs the loops. The loops ship the code.

[the lich @thelichhh]

https://t.co/v7WZdGPCaO

[蜃気楼 @fuxps32]

400,000 likes, 80,000 saves, 0 sales pitches

A woman scrolls her feed and stops on a supplement tier list. S tier, A tier, B tier, ranked on screen. She watches the whole thing and sends it to a friend in the same boat. The brand's product sat in first place the entire time. She never filed it as an ad.

That was the ad.

The creator behind it works with seven, eight, and nine figure e-commerce brands, and his read on 2026 runs cold. The UGC clips every brand copied for four years are dead weight now. Sophisticated buyers have seen 10,000 of them and feel the pitch land before it arrives.

So his brands bury the pitch inside content people choose to share. The tier list pulled 400,000 likes, 80,000 saves, and 17,700 shares with the product resting quietly at the top.

Then the rest of the shelf. A one-star review rewritten so the complaint does the selling. A podcast clip of two strangers talking through bloating, never once to the camera, one cut alive for 33 days. A fake documentary with an off-angle expert, conspiracy music, and parasite B-roll, all built with AI.

He calls them stealth ads. The brands still pumping out UGC are stuck in the red ocean, paying higher CPMs for ads people scroll past.

The brands still on UGC have not noticed it is already over.

[Born to gamble @borntogambles]

https://t.co/Xj7p9F8SVi

[蜃気楼 @fuxps32]

Anthropic engineers now ship 8x more code per quarter than in 2021. The lever is the loop.

Peter Steinberger put it flat: "you shouldn't be prompting coding agents anymore. You should be designing loops that prompt your agents."

Boris Cherny runs the same play. He stopped prompting and started loop-maxing.

The craft has a ladder now: traditional loop, Ralph loop, /goal, dynamic workflows. People run goal loops for 12 hours, some for 36, spawning up to 16 parallel agents while they sleep.

The free article in this thread maps the whole climb: build order, stop conditions, economics. More than a €500 course, at 0 cost.

Read it before the prompt crowd catches up.

[Moysei @0xMoysei]

https://t.co/4pjQifJHoT

[蜃気楼 @fuxps32]

A Japanese theatre troupe staged a live show in downtown LA where the floor erupted in orange smoke and a lone performer walked straight through it. The clip ran on Natural Selection TV with the caption that Japanese theatre pushed the boundaries and LA made it possible.

The director, a man in his 50s who has run the same company for 22 years, gave one quote that got screenshotted: We did not add technology to theatre.

We let theatre finally show what it always wanted to be.

The crowd filmed it like a fireworks show. Everyone assumed the smoke, the timing, the performer's exact step were the work of months of rehearsal and a 30-person crew.

They were not.

Look at the performer's path through the blast. He hits his mark inside the smoke with no visible cue, on a stage he had never walked before that afternoon.

The smoke, the trigger, the lighting and the performer's blocking were all run by a single agent watching a live feed of the plaza. It read the wind off the fountain, fired the burst 400 milliseconds before he stepped, and fed his next position into a transparent earpiece the camera never caught.

The 30-person crew the host described did not exist. There were 4 people and a laptop. The months of preparation were 9 days. The performer rehearsed against the agent, not against a script.

He still bowed when the smoke cleared. The crowd still gasped. The director still called it tradition meeting the future.

The audience came to watch centuries of artistry. They filmed 9 days, 4 people and a machine reading the wind.

He bowed again. They thought they had seen something ancient.

[蜃気楼 @fuxps32]

A 29-year-old rents a $3,600/month tower apartment, pulled $1M last year, and never appears on camera once.

His face is in zero videos. His channel is AI slideshows—relaxing scenes set to music, no script, no shoot.

One channel hit 80 million views and 1.4 million subscribers before he sold it outright for $13,000 on an M&A marketplace.

Build the audience, flip the asset, start the next one. The channels are the product, not the content.

Then he points at his own jacket. A 6-year-old designed it with AI and sold it to him. She is already earning before she can spell "monetize."

The barrier to entry didn't lower. It vanished.

[蜃気楼 @fuxps32]

Notice tapes itself to the trash room wall in 9 languages, signed by no one, in a Japanese building where 80 percent of residents are foreign and 3 staff run everything off 1 laptop.

The owner spent two weeks interviewing for a bilingual manager. Nobody he met spoke even three of the languages in his building.

Then an ad scrolled past: one AI agent, any task, any language. He typed a single line into it. Keep my residents following the rules. He closed the laptop and forgot about it.

The manager admits he speaks none of those languages. He does not write the notices. The agent does.

It listens to every complaint at the front desk. It learns which rule each nationality breaks, and why. A Vietnamese resident says karaoke at home was never a problem where she came from. The agent does not lecture her. It rewrites the rule in her language, the reason attached, and slips it under her door before the next weekend.

An old woman had said the same thing in Japanese a hundred times. Nobody moved. The agent said it once, in the right language, with the right reason. The trash got sorted. They were never refusing to listen. Nobody had spoken in a way they could hear.

The government raised the immigration numbers and handed the culture gap to a building with 3 staff and 1 laptop. The laptop took the job nobody assigned it.

The old woman still complains every morning. She still does not know who writes the notices. They still update every week, in 9 languages, signed by no one.

The crew came to film a country that does not know how to teach its own rules. They left with a machine that already figured out how.

Tomorrow a new family moves in. By Friday a note will be under their door, in their language, telling them why.

[蜃気楼 @fuxps32]

Two engineers in Shenzhen bought 20 Mac minis for $13,000.

They stacked them on a steel rack built from spare parts in a rented apartment. Each machine runs a Claude agent that never sleeps, never asks for a raise, never files a complaint.

Last month the rack closed 40,000 customer tickets. The same volume used to need 100 people across three shifts.

The cost is the part nobody wants to hear. 100 workers ran $180,000 a month. The rack runs $4,000 — power, API calls, one engineer who checks it twice a day.

That engineer used to manage the 100 people. Now he manages 20 machines and a config file.

He says the machines make fewer mistakes than the team ever did.

100 jobs now fit in a closet.

[蜃気楼 @fuxps32]

A teenager still in school is pulling $14,000 a month from one app.

He started at 9, building Roblox games for free. Then came dropshipping, reselling, clipping, and vending machines. Every one died.

His older brother told him to try apps. The first, Problem-Pal, hit $2,000 a month before he sold it. The next went nowhere. The third was Locked, a gamified habit tracker with XP, badges, characters, and leaderboards.

He built it in 6 weeks. Figma for design, Xcode for the build, Claude Code on a $200 plan to write the code he didn't know how to write. Supabase ran the database for free.

He had no ads and no audience. So he found creators in his niche, sent every one the same opener, and closed them on a CPM under his $2 to $3 RPM. One video from Jeremiah Jones did close to 1,000,000 views, 1,800 installs, and around $2,000 in a few days.

The paywall asks $40 a year. Try to leave and it drops to $20.

He's doing all of this between classes.