real setups, live dashboards. dm open.
Joined October 2026
- Tweets103
- Following39
- Followers370
- Likes64
Walrus retweeted
GTA 6 isn't out yet. Anthropic's new Fable 5.2 already built me one
one prompt. I was not ready for how close it got to the real thing
an indie studio quotes $90,000 and 8 months for a playable prototype like this. Fable 5.2 did it in 2 hours 14 minutes on a $200 plan
what's in there:
→ on foot: sprint, jump, kick, fists, a pistol with real ammo and reloads
→ crime: a witness calls the cops, the stars stack up to 5, police shoot back
→ cars: jack any car, 3 gears, handbrake, headlights, live traffic and crashes
→ garage: 10 paint jobs, bumpers, 5 wheel sets, rim colors and neon underglow, all paid from your balance
→ missions: a 9-checkpoint race on a timer and a courier run that pays $600
→ the city: 9 districts, sunny day into neon night, a minimap, hidden cash that glows green and a pelican on the pier
even the sound is code. every gunshot, engine and siren is generated in python, zero audio files
the prompt:
"you're a senior game dev. build a third-person open-world crime game in a sunny 80s miami-style city. make the loop tight: walk, fight, steal a car, get chased, earn money, spend it. every system feeds the next: crimes bring heat, heat brings cops, missions pay, money unlocks the garage. cars need weight: gears, skids, real crashes. the city has to feel alive with traffic, people and day into night. no audio files, generate every sound in code. after each system, play it yourself, screenshot it and fix what looks wrong. don't stop until a stranger could play for 10 minutes without hitting a bug"
here's what it built ↓
this is f*cking huge... Google just shared 303 free skills that turn Claude into your assistant for Gmail, Calendar, Drive and Sheets
one works like an executive assistant. one like a sales ops lead. one like an HR coordinator. one designs a full app screen from one sentence
64,000 GitHub stars between them. an executive assistant costs $4,350 a month. here all 303 cost $0
how to set it up:
1. open Claude Code
2. ask it to install the skills from the links below
3. connect your Google account
4. tell it: "sort my inbox and find me one free hour tomorrow"
workspace: github.com/googleworkspace/c…
ads and cloud: github.com/google/skills
design: github.com/google-labs-code/…
one week in, your inbox is sorted, every receipt lives in one sheet and Friday ends with a digest of everything that moved
64,000 developers already set this up. the full workflow is below ↓
this is f*cking unreal... the CEO of Obsidian just shared 285 plugins for motion design for free. ALL OF THEM
one turns a rough sketch into an animated scene. one flies a camera across your canvas. one makes slides morph like Keynote. one gives you an endless whiteboard for storyboards
30 plugins for slides that move
222 plugins for canvas, flowcharts and presentations
33 plugins for drawing and sketching with Excalidraw
9M downloads on Excalidraw alone. designers charge $1,500-4,800 for one animated presentation. here all 285 cost $0
how to grab one:
1. open Obsidian → Settings → Community plugins
2. hit Browse and search the name
3. install and enable
4. open a new note and build your first scene
slides: obsidian.md/plugins?search=s…
canvas: obsidian.md/plugins?search=c…
drawing: obsidian.md/plugins?search=e…
start small: one scene today for the idea you explain most
all 285 plugins on one map below ↓
IT IS BECOMING FUCKING DANGEROUS
A solo builder turned a plain folder of Markdown files into a living second brain and watched the graph light up the moment Claude loaded the map.
CLAUDE CODE / SECOND BRAIN
What the graph shows:
- CLAUDE.md / LOADED - map of the brain · under 200 lines · every session
- rules/people.md / MATCHED - people/ loads only on match
- skills/weekly-review / RUN - review projects · flag stale people
- memory.md / SAVED - always put a source under every claim
- rules/writing.md / CONFLICT - contradicts CLAUDE.md naming rule · fix proposed
- 3,211 → 3,365 records · 11,500+ links · one topic per file
The article said the second brain is just folders + a CLAUDE.md map. No Obsidian. No database. Just text the agent can read. This is that map running live - Claude loads first, rules fire only on match, skills run when called, and every claim keeps its source.
Most people still dump notes into apps the agent can’t see. This one keeps everything flat so the graph can load, match, run, save, and flag conflicts before anything goes wrong. The operators are nervous because the folder now catches its own mistakes.
The map is loaded. The graph is watching.
It is becoming f*cking replaced
Dario Amodei said the best chief of staff you’ll ever have will never ask for a raise, never miss a promise, and will ask you before it touches anything.
A solo builder just turned a flat folder of Markdown into that person and let Claude Duet spin the memory graph while every node asked first.
CLAUDE DUET · CLAUDE CODE · OPUS 5.5
CHIEF OF STAFF · MEMORY GRAPH · ASKS FIRST
What the graph shows:
- MEMORY, SCHEDULER, DEPLOY, RETRY, FIX, MERGE lighting in clusters
- AGENT 07 / AGENT 12 spinning, CLUSTER 03 → 17 growing
- PLANNER, RETRIEVAL, TOOL CALL, CODE EXEC, SAFETY live
- RULES 4 · ALLOW 4 · ASK 4 · DENY 2
- SENT W/O YES 0 · BRIEF 07:00 · REVIEW FRI 16:00
- PROACTIVE · LID CLOSED OK
The article said Claude Duet comes to you with what needs attention, finishes the work, keeps going when you close the laptop, and asks before it does anything in your name. This is that loop as a sphere network.
Most people still pay a chief of staff $8k a month. This free local graph replaced him, never misses a promise, and never sends without asking.
The map is loaded. The graph is asking.
It is becoming f*cking abyssal
A solo builder just turned a flat folder of Markdown into a living octopus and dropped it 400 meters deep so Claude Code could grasp every file while the human stayed dry.
CLAUDE CODE / SECOND BRAIN
// ABYSS-8 · 8 ARMS · 2,912 GLYPHS
What the abyss shows:
- octopus rising, 8 arms trailing, CLAUDE.md locked first
- sonar pings, jet drive 124 BPM, ink tank full
- arms grasping: memory.md, rules/people.md, agents/review.md, mcp/servers.json
- CLAUDE.md branched, tree graph live, commits lighting the tentacles
- memory.md absorbed into core, 380 glyphs, 100%
- fan-out - 8 arms, every file hot
- ink render: SECOND BRAIN written in 2,600 glyphs
The article said the second brain is just folders + a CLAUDE.md map. No Obsidian. No database. Just text the agent can read. This is that map running as an abyss - Claude dives the index, the arms seize the files, and the human still owns every claim.
Most people still pay a knowledge manager $3,000 a month to keep their notes in order. This one replaced him with a free local octopus that never forgets a source, never invents a link, and works while you sleep. The specialist is now scared because the folder does his job better.
The map is loaded. The octopus is diving.
It is becoming f*cking sviny
A 19-year-old dev from China just turned a flat folder of Markdown into a living world and dropped pink pig agents into the SVINY TERMINAL so they could fly the map while Claude Code expanded the rules.
CLAUDE CODE / SECOND BRAIN
SVINY TERMINAL · 3,276 → 3,456 NODES
What the terminal shows:
- CLAUDE.md loads first, 142 lines, auto memory on
- Flying pink pigs circling the village nodes
- rules/people.md matched → expanding → exploded into atoms
- One topic / file, sources under claim, paths and preferences visible
- Molecule diff and link matrix live under the houses
- Reassembled, memory saved, sources under every claim
- Spin audit 128/128, globe index rotating
The article said the second brain is just folders + a CLAUDE.md map. No Obsidian. No database. Just text the agent can read. This is that map running as a world - Claude loads the index, the pigs fly the links, and the human still owns every claim.
Most people still dump notes into apps the agent can’t see. This 19-year-old keeps everything flat so the terminal can light the nodes, cite the file, and never invent a source.
The map is loaded. The pigs are flying.
11
28
1,176
Walrus retweeted
Are you doing your f#cking job or training your replacement?
Meta started recording where employees click, what they type and what they open on screen. To teach AI to do the same things. You're still collecting a paycheck while the machine learns to work without you.
In April, Meta rolled out MCI on US employees’ work computers: mouse movements, keystrokes and occasional screenshots. Meta confirmed the program to Reuters.
Open a menu. Complete a task. Fix a mistake. Your working day becomes a training manual.
Meta’s CTO described the goal plainly: agents do most of the work, humans direct and review. How many humans that requires is another question.
The company promised to protect sensitive information. The data was for training models, not evaluating employees.
After employees pushed back, Meta added a pause button. For 30 minutes at a time.
Then the sh#t got worse.
According to documents reviewed by Reuters, sensitive data was accessible to all Meta employees. Private conversations. Full AI prompts. Personnel information.
One employee said they had accessed medical and tax documents on their work computer. They had been promised protection too.
On June 22, Meta announced it was pausing the program. The company said it had found no indication of improper access at that point.
But the arrangement is clear: the company pays you to work while teaching a machine to perform your tasks.
When your boss asks you to “help implement AI,” will you understand what you're helping build? Or only when you're told to get the f#ck out?
🚨WHAT THE F#CK IS GOING ON?
Sam Altman thinks scarce memory is being wasted in your f#cking computer. It would work more efficiently in the cloud. A major manufacturer has already announced it is dropping its consumer brand to prioritize bigger customers.
The argument is about efficiency. For buyers, the consequence is already concrete: one less familiar memory brand.
In the video below, Altman explains why scarce RAM should get maximum use in data centers rather than sit in personal PCs.
Start a year earlier, when Samsung announced a planned partnership with OpenAI.
In October 2025, Samsung put OpenAI's projected demand at up to 900,000 memory-chip wafers per month. A forecast, not a purchase. But an enormous appetite.
Wafers are the silicon discs used to manufacture chips. Then came the announcement for consumers.
Two months later, Micron announced its exit from the Crucial consumer business.
The company explicitly cited growing demand from AI data centers. Dropping Crucial would help supply larger strategic customers in faster-growing markets. The brand had lasted 29 years.
Consumer shipments were scheduled to end in February 2026. Warranty support would remain. Millions of buyers got a thank-you. Bigger customers got priority.
Then, in October 2026, Altman called keeping scarce memory in personal PCs "quite terrible" from a global allocation perspective. His estimate: it gets used 5% of the time.
Your choice isn't in that calculation.
You bought a computer to have it available when YOU need it. Not to keep it constantly busy.
Local AI already offers an alternative to someone else's prices and rules. But it still needs hardware.
If buying your own becomes harder, how much f#cking choice will you have left?
It is becoming f*cking spherical
A solo builder just turned a flat folder of Markdown into a living sphere and let Claude Code explode the rules until every claim had a source.
CLAUDE CODE / SECOND BRAIN
GRAPH · 3,276 → 3,456 NODES · 11,714 → 12,134 LINKS
What the sphere shows:
- CLAUDE.md loads first, 142 lines, auto memory on
- Nodes lighting: inbox.md, people/, projects/, notes/tax-policy.md, rules/writing.md
- rules/people.md matched → expanding → exploded into atoms
- One topic / file, sources under claim, paths and preferences visible
- Molecule diff and link matrix live
- Reassembled, memory saved, sources under every claim
- Spin audit 128/128, globe index rotating
The article said the second brain is just folders + a CLAUDE.md map. No Obsidian. No database. Just text the agent can read. This is that map running as a sphere - Claude loads the index, expands the rules, and the human still owns every claim.
Most people still dump notes into apps the agent can’t see. This one keeps everything flat so the sphere can light the links, cite the file, and never invent a source.
The map is loaded. The sphere is spinning.
Holy sh*t! Claude vs GPT vs Gemini vs DeepSeek
Same prompt. Same fight.
Price: DeepSeek (40x cheaper)
Brains: Claude
Honesty: Gemini
What people actually prefer: Gemini
Special skills: GPT-6 Astra
Nobody sweeps. Every fan account is screaming their model is the best.
The independent numbers say the scoreboard is closer than any of them want to admit.
Why is the marketing always so much louder than the actual gap?
10
35
2,567
It is becoming f*cking harvested
A solo builder just locked Claude Code + Opus 5.5 into a live lab and dropped JEV as the router that decides which drone grabs the next chunk.
CLAUDE CODE · OPUS 5.5
JEV · FLEET
core 98% · 2,400 RPM
What the lab shows:
- Central vortex spinning, agents 4/4 online
- Drones M-01 / M-02 / M-03 + X-01 harvester scanning, seeking, grabbing
- JEV gate sealing chunks, 0/8 → 8/8, status harvesting
- Node graph live, latency dropping
- Code in jev_router.ts: harvest, seal, threshold, upload
- Views 3,790,000 · prompts 9 · skill free · Claude
- Machine specs: V8 4-stroke 2,400 RPM, sandworm layers, HM-4 limbs
The split is the same. Claude writes the prompts and the code. JEV decides which frame is worth the grab. The drones only move after the gate says yes.
Most people still run one chat and hope. This one has a fleet, a router, and a core that doesn’t stop until every chunk is sealed.
The lab is live. The harvester is moving.
9
26
668
It is becoming f*cking cellular
A solo builder just turned Claude Duet into a living plant cell and cut it open so you can watch the chief of staff work from the inside.
CLAUDE DUET · CHIEF OF STAFF
A PROACTIVE-AGENT CUTAWAY
memory · brief · asks first
What the cell shows:
- Wall holding 0.62 MPa turgor, cellulose locking the structure
- Nucleus + nucleolus streaming DNA → mRNA → ribosomes
- Mitochondria and chloroplasts running ATP and photosynthesis live
- Golgi and ER shipping vesicles, vacuole holding 90% of the volume
- Full X-ray cut: every organelle in place, focus lock 98%, asks first
The article said Claude Duet is a proactive agent that comes to you, keeps working when you step away, and checks before it acts in your name.
This is that agent rendered as a cell 0 memory in the nucleus, brief in the chloroplast, permissions in the wall.
Most people are still waiting on the research preview. This one already has the organelles, the map, and the rule that nothing leaves the cell without a yes.
It is becoming f*cking routed
A solo builder just turned a flat folder of Markdown into a living decision globe and dropped JEV as the halo kernel that never lets the LLM answer yes or no.
CLAUDE.md loads first. One file, one topic. The agent reads the map. JEV sits in the cortex deciding what moves.
What the globe shows:
- 01 MD MAP - CLAUDE.md loads, 22 walls live, decisions climbing past 9k
- 02 INBOUND LEAD - score lead_quality high 0.89, book_call, calendar link sent
- 03 EMAIL TRIAGE - phishing 0.97, quarantine, LLM never woke
- 04 YOUTUBE - worth_reply yes 0.91, draft reply, only then the writer wakes
- 05 BROWSER AGENT - action chosen, hand action off, code clicks checked
- 06 X FEED - viral_fit 0.92, draft_short, you still hit Post
- 07 SUPPORT - urgency 0.71 below 0.85, HOLD, human gate
- 08 THE SPLIT - LLM writes, JEV decides, code acts, $0.42 for 10k, ~256 ms
- 09 GUARDRAILS - spend cap, actions limit, never auto-post to X, progress saved
The article said the second brain is just folders + a CLAUDE.md map. No Obsidian. No database. Just text the agent can read. This is that map running as a decision globe - and JEV is the halo that keeps the ring spinning, seating the choice, tracing the confidence.
Most people still burn frontier tokens on “is this spam?” while the map sits empty. This one keeps everything flat so Claude can load the file, JEV can gate the fork, and the human still owns the irreversible step.
The map is loaded. The globe is live.
It is becoming f*cking operational
This system is already on the roads.
Every car, truck and bus is getting its own file in real time. No app. No database. Just flat Markdown and a CLAUDE.md map that tells the agent exactly where to look.
Anthropic’s people keep talking about agents that can actually use your files. This is what it looks like when the rule is enforced at highway speed:
CLAUDE / SECOND BRAIN
every object gets its own file
- YOLO + ByteTrack running live
- 20–26 vehicles tracked per frame
- Each one written as its own note: #21 TRUCK 63 km/h, #331 CAR 100 km/h, #374 BUS 101 km/h
- ONE TOPIC = ONE FILE
- Agent opens the exact file and cites the number
- Speed calculated from pixels → metres → km/h, smoothed, saved
No shared notes.md. No “the agent will figure it out.”
If the file doesn’t exist, it fails instead of inventing a speed.
Most people are still waiting for some research preview that can read a folder.
This one is already citing individual vehicles on the highway while the rest of the timeline is arguing about prompts.
The guide said the second brain is just folders and a map.
The roads are already running it.
Claude Second Brain & Folder Workflow: The Complete Guide.
It is becoming f*cking profitable
6 billion got stolen from Binance. For most it was a disaster. For the people who already had a second brain wired to the desk it turned into one of the cleanest nights of the year.
While the rest of the timeline was panic-refreshing, the atlas kept running:
CLAUDE / SECOND BRAIN
Markdown · Map · Memory · one topic per file
- Order-flow plane spinning on BTC-PERP / ETH-PERP / SOL-PERP
- Matching core: 160-180 fills/s, p50 under 4 ms
- Arb bridge ETH↔SOL live, 2.8-4 bps, basis locked
- CLAUDE.md loaded: Jev scores the edge, gate at 0.85, router places, book hedges
- Fills/min 2,100-2,700
- Net PnL sitting +$10k to +$15k
- Win rate 60-62%
- Every tick still runs through the same flat map
The guide said the second brain is just folders and a CLAUDE.md. No Obsidian. No database. Just text the agent can read and cite. This is that map turned into a trading cortex - Jev decides the fork, Claude keeps the memory, the desk only acts when the confidence clears.
Most people still dump notes into apps the agent can’t see and then wonder why they missed the move. The ones who kept everything flat already had the edge scored, the hedge ready, and the position sized before the headlines hit.
6 billion left. The atlas stayed on.
BREAKING: CZ (founder of Binance) says it looks like Binance got hacked. more than $2.5 billion left its hot wallets, and Tether and Circle froze $1.8 billion of it within hours
hacks this size rarely start with code. the $1.5 billion Bybit hack last year started on one developer's laptop. one machine, one process nobody noticed
so last night I had Claude Code check mine. it found 18,742 webcam photos and video clips on my MacBook that I never knowingly recorded. the oldest folder was 14 months old. the newest was from that morning
I opened the previews expecting forgotten video calls. instead I saw my desk after midnight, friends walking through the room and an empty chair while a conversation continued in the kitchen
the files were buried in an application's storage folder. each had a timestamp, and the videos had audio. none appeared in Photos or the folder where I normally saved recordings
I made a backup and had Claude inspect the copy. it grouped the files by date, checked their metadata and traced the directory to a background capture app configured to launch when I logged in
then I checked the original folder. the latest clip had been created 12 seconds earlier. it showed me leaning toward the screen, opening the same files I was now reviewing
Claude checked the running processes and identified the recorder. I stopped it and watched the folder for another 10 minutes. no new files appeared. the final clip ended with me reaching for the keyboard to shut it down
You can't stop the next exchange hack. you can make sure it doesn't start on your laptop. the prompt that found the recorder on mine is in the first reply
It is becoming f*cking indexed
A solo builder just turned a flat folder of Markdown into a live second brain and the agent is already walking the dependencies.
48k files. 10 folders. One map. Claude loads CLAUDE.md, scans the graph, and holds every write behind a guard.
What the lab shows:
- CLAUDE / SECOND BRAIN · one topic per file
- Trace stream opening index.sql, editing candidates, queuing checks
- Agent lanes: PLAN · REWRITE · COMPARE · CHECK - all READY
- Query editor rewriting SELECT SUM, CREATE INDEX, EXPLAIN in real time
- Dependency walk on Q03 index candidate
- Check runner holding PLAN / READ / WRITE until the candidate clears
- Write guard watching incoming writes, review gate pending
- C8 / C3 scanning · 7,183 leaves / links attached
The guide said the second brain is just folders + a CLAUDE.md map. No Obsidian. No database. Just text the agent can read.
This is that map running as a cortex - the agent finds the file, cites it, and never writes without the gate.
Most people still dump notes into apps the agent can’t see. This one keeps everything flat so Claude can load the map, walk the links, and keep the irreversible step for the human.
The map is loaded. The graph is live. The write is still waiting.
It is becoming f*cking proactive
A solo builder just stood up a Claude Duet chief desk that scans Polymarket while the laptop is shut.
200 wallets. 48k trades. Filters fire. Clusters form. Fresh wallets get pinged before the result.
What the desk shows:
- Memory · inbox scan · brief 07:00
- Fund scan top 200, then split 20, then analysts read in parallel
- Filters: age ≥ 60d, events > 2 → wallets drop in real time
- Risk review: pnl < $500k, public news, team, company → DROP or KEEP
- ALERT 01 · CLUSTER: BRA-PRES-26, 2 fresh wallets → 1 market, age 26d / 48d
- New wallet age 0d, first bet FED DEC · CUT $48.2k → PING owner before result
- Dispatch across 20 workers, funnel scanned 200, risk keep 2, alerts 02
The course said a proactive agent comes to you with what needs attention, finishes the work, and asks before it acts in your name. This is that loop on prediction markets: Claude reads the board, gates the flags, surfaces the cluster, and only pings when the pattern is real.
Most people still refresh leaderboards by hand. This one keeps the scan running, drops the noise, and holds the irreversible step for the human.
The desk is live. The brief is already waiting.
It is becoming f*cking mapped
A solo builder just turned a flat folder of Markdown into a living second brain and dropped JEV as the halo kernel inside it.
CLAUDE.md loads first. One file, one topic. The agent reads the map. JEV sits in the cortex deciding what matters.
What the brain shows:
- 01 CLAUDE.MD - the map loads first, spin 2.63 rpm
- 02 NOTES/ - one file · one topic, dies seating, 0 rejects
- 03 CORTEX - map points, agent reads, pulse 10- bpm, fibres lighting
- 04 @-LINKS - MEMORY.md stays in focus, 1128 links, every bay hears
- 05 PEOPLE/ - one card · one person, echo rings
- 06 PROJECTS/ - where you paused · next step, underside conduit live
- 07 RECALL - find the file · cite it, probe PASS
- 08 MEMORY - learnings · decisions, bays locking
- 09 FULL BRAIN - flat .md · no apps, 384 bays, 1504 lines live, full halo
The article said the second brain is just folders + a CLAUDE.md map. No Obsidian. No database. Just text the agent can read. This is that map running as a cortex - and JEV is the halo that keeps the ring spinning, seating the dies, tracing the links.
Most people still dump notes into apps the agent can’t see. This one keeps everything flat so Claude can load the map, JEV can gate the recall, and the human still owns the irreversible steps.
The map is loaded. The halo is live.
It is becoming f*cking orchestrated
A solo builder just fused JEV as the decision orrery with Claude Code + Opus 5.5 as the worker.
JEV gates. Claude migrates. Nothing irreversible runs without the ask.
What the orrery shows:
- Live context orbit: Repo Satellite, Docs Balloon, Web Probe, Terminal Drone, MCP Relay, Paper Orbiter
- Tasks flowing: MIGRATE PROD DB → HUMAN GATE (irreversible, ask before run) → dry-run ok → skipped
- MIGRATE API V1→V2: 79 files, 9 steps, +530/-517 diff, risk med 0.42 → gate auto 0.96 PASS
- TRIAGE 40 ISSUES: dupes ×8, 4 buckets, labels ×37, 2 need human → gate auto 0.95 PASS
- 12 subagents dispatching in parallel: read, grep, plan, edit, bash, review, web, gate
- Tool calls 19,500+, latency ~250 ms, shared tree, cache hit 90%+
The article said the split is the whole product: if it creates → LLM, if it picks/scores/gates → JEV, if it does the thing → code. This is that split running as an orrery. Claude Opus plans and edits. JEV decides whether the next step is safe. The human gate sits on anything irreversible.
Most builders still mash every yes/no into one expensive call. This one keeps the expensive model on the work and the cheap one on the fork.
The second brain finally has an orbit.
19
49
1,441