Atlas

THIS IS INSANE. ONE ENGINEER SHIPPED 1,000 PRs IN A MONTH WITH AI AGENTS 😱 A SpaceXAI engineer who builds Grok Bot showed her setup in a free 57-minute workshop. She woke up to 20 PRs her agents had already merged. 3,000+ PRs in 5 months. 3:27-5:14 - agents auto-merge PRs while she sleeps 5:46-6:40 - the #1 skill: the agent runs the app and checks its own work 14:50-17:22 - every agent mistake becomes a skill 39:08-40:58 - CI fails on the patterns agents get wrong Paid agent courses cover less than this. One person running a software company isn't a joke anymore 🤣 Bookmark this

20,000 PEOPLE PREORDERED A $399 ROBOT DUCK. ITS WHOLE BRAIN IS FREE ON GITHUB That's hugging face's ceo's number. The first batch is already sold out Watch jensen at gtc berlin. He asks the ducks to nap. Three drop on the spot. One keeps walking a few steps, then lies down with the others Nobody animated that. Every move is a neural policy trained in simulation, thousands of ducks at once, then loaded onto the real one What $399 gets you: > 25 cm, under 800 g, 15 servos > A small rockchip board running a 50 hz control loop > Walks, roller-skates, picks things up, stands back up after a fall > 7 behaviors out of the box > Apache-2.0 code, so you can train your own moves 20,000 x $399 is almost $8m in preorders. For a duck First ones ship before christmas

OPUS 5.5 BUILT THIS GAME IN 79 MINUTES. KEEPING PLAYERS IS THE HARD PART Day 1 was two agent runs, 46 and 33 minutes. 3,452 lines and a floating island that tilts under every house you place. Day 2 took four more runs and $45.22. 14,918 lines, five chapters, a daily island, one shared world island. Then real players showed up. 27% quit on the first island, median 24 seconds. 15% beat chapter 1. Most runs end around second 170. CrazyGames expects 80% of players to stay past one minute. Skyhold kept 36%. The code worked. The first island ran on a timer, and houses slid off before anyone learned the tilt. So the next runs went into the first three minutes: a calm first island with no timer, 30 short levels between the islands, 12 languages at 1,226 strings each. A separate Opus pass audited the server and found 13 holes in 5 minutes for $2.19. Free in your browser, phone too: skyhold.lol

ANTHROPIC'S PROMPTING WORKSHOP IS GOING AROUND AGAIN. HERE IS WHERE EVERYTHING IS Two engineers from Anthropic's applied AI team take one prompt from wrong to production in 24 minutes. The task: read a hand-filled Swedish car accident form and a sketch, then decide who was at fault 0:00 → 2:58 - the bare prompt. Claude decides it was a skiing accident 2:58 → 6:49 - the structure: task first, then content, examples, key rules repeated at the end 6:49 → 9:25 - told to stay factual, it stops guessing and won't name a driver 9:25 → 13:11 - the form never changes, so it moves into the system prompt and gets cached 13:11 → 17:31 - few-shot examples, images passed as base64 17:31 → 21:49 - reading order: form first, sketch second 21:49 → 23:32 - verdict wrapped in XML tags, prefill for JSON 23:32 → 24:52 - extended thinking as a way to debug the prompt Version one called it a ski crash. The last version names vehicle B inside a tag a program can parse Full workshop below

Someone who isn't a trained developer used Claude Code to build a pizzeria its own ordering app, so regulars stop paying DoorDash 25-30% > Claude Code wrote the app, the backend, Stripe payouts straight to the restaurant and the hookup to its Clover register > it also wrote the tests that catch mistakes before deploy > the restaurant pays $99 a month and 0% per order What Claude couldn't do: get the first owner to say yes. That took walking in with a phone and showing him his own menu in the app Anthony's Pizzeria in Allentown, PA is live now

Black October started with one bug: AI found a critical hole in PHP, the software behind an estimated 70% of websites. > OpenAI says its models found it, then spent two months building fixes with the maintainers before disclosure > the NSA and the FBI told every organization running PHP to get ready to patch > on air the warning was blunt: "now with AI, you can actually reverse engineer the patches into an exploit themselves" The fix is out. The time between a patch and a working exploit is what just got shorter.

Claude Opus 5.5 built an F1 game with 20 cars, start lights and DRS. No game engine, no 3D assets, one session. Then it tested the game itself: > held the throttle: 0 to 186 km/h in 6 seconds > steered, braked, hit a wall, hit reset > ran a full race at 6x speed: 20 of 20 cars finished, 60 fps > looked at its own screenshots and fixed the washed-out paint and the single-file AI The prompt ended with "Start building, then verify it, and tell me when it's done." The full course on building and shipping a browser game like this is below ↓

Claude Duet, Anthropic's new proactive agent. 17 minutes of real use from someone with early access: 0:48 → it's already connected to your Claude account, nothing to set up again 4:46 → why it's the most proactive agent in the test: a check-in task every 3 hours, 9am to 10pm 6:29 → type / for hidden skills. /morning renders your day as a styled brief 7:25 → it read 33 emails, 47 DMs and 11 docs, then wrote a guide to how you write 9:03 → one line, no fancy prompt: a year of YouTube data turned into an insights dashboard 12:08 → where it breaks: an Uber booking stuck on a "verify you're human" check 13:55 → how it stacks up against Grok Bot, Muse and ChatGPT Dots Waitlist only for now: Pro and Max, US.

This is what the $120-an-hour IT guy looks like when it's Claude. One prompt: find every duplicate on this drive, build me a report, delete nothing yet. > 17.85 GB of duplicates found > an interactive worklist: keep oldest, keep newest, sort by wasted space > it re-checked 5 random groups before showing the report > the delete script moves files to the Trash, not rm -rf After two weeks of cleanup: 50 GB back, hundreds of duplicates gone. Cleaner apps do part of the job. This one asks how you want it organized first.

Someone open-sourced the kit that put real Minecraft inside GTA V, and this is a f*cking goldmine Point Claude Code at any PC game you own, describe the mod, it does the rest: → Finds the game and works out the engine → Reads the real code and builds the mod → Makes the art, 3D and sound → Backs up your saves first → Tests it in the running game → Writes down what it learned for the next agent One test run put Black Ops 3 zombies inside GTA V: ~4,000 lines of C++, 47 hot reloads, 60 screenshots, all by the agent 8,482 stars in 3 weeks. Two commands to install: github.com/rehan-remade/univ… The article below is the other half: building and shipping your own game with Opus 5.5 ↓

The trend of the weekend: people are vibe coding their TVs One prompt to Claude Code and the ad-filled home screen is gone. No root. Nothing uninstalled On this Android TV free memory went from about 370 MB to 600-650. Roughly 90% more room The prompt is on GitHub. 2,692 stars in 2 days Next up: someone already jailbroke their Kindle the same way

Anthropic's own team put their small business skills in the plugin repo, and this is a f*cking goldmine you install it once, connect your tools and claude runs your back office here are 7 worth looking at: skill 1 → invoice chase drafts reminders to customers who pay late skill 2 → pay the bills codes every bill and stages the payment run for your approval skill 3 → cash flow snapshot tells you if you'll make payroll skill 4 → speed to lead answers every inbound lead overnight skill 5 → hiring screener turns a pile of applications into a shortlist skill 6 → reactivate finds customers who stopped buying and drafts the win-back skill 7 → restock drafts the purchase order before you run out worth installing before you hire your first assistant

Claude Motion one-shot a 30-second real estate explainer. Then one line changed it: "Change the median price to $640,000 and make it nighttime." It took 45 seconds. The median now counts up to $640,000, and the neighborhood is at night: moonlit houses, streetlamps at each intersection, homes glowing amber as they sell. Claude also said what it hadn't changed. The floating price tags were "still placeholders clustered around the old $612K figure," and it offered to move them to around $640K. That's step 08 of the article below: ask for edits in numbers, not adjectives. The other 10 steps are there too ↓ polymarket.com/event/which-c…

762 people played a game built with Opus 5.5 in its first day or so. The average session ran 1 minute 41 seconds, about long enough to reach the end of the level. The game is Tidebreaker, a 2.5D platformer with a boss chase. Its maker says the puzzles, the graphics and the game logic were all made with AI, with Claude Code on Opus 5.5 doing most of the work. > PLAYERS - 762 monthly, 634 daily on the dashboard > SESSION - 1 min 41 sec on average > STACK - Claude Code, Blender, Tripo for 3D, ElevenLabs for audio The article below walks through building a mobile game like this in under 5 hours and getting paid for it. The model that built it is still second in the leaderboard race. Polymarket has Google at 70% to hold the top AI model at the end of October, Anthropic at 28.5%. polymarket.com/event/which-c…

Someone gave Claude Motion 90 days of heart-rate data and one paragraph. This is what came back. The file was hrv_recovery_log.csv. The prompt asked for a 30-second explainer, white background, very little text, and set two rules: "Calculate every number from the CSV" and "Show relationships without claiming causation." > SLEEP - "a positive impact on recovery" > ALCOHOL - hurt recovery "by quite a bit" > LATE MEALS - 70.2 ms without, 58.1 with, -12.1 ms > WORKOUTS - "some impact on recovery, but not a ton" The takeaway after one watch: "get plenty of sleep and don't drink too much." The model picker under the prompt reads Opus 5.5, Medium. Polymarket has the next Opus out by November 30 at 79%, by October 31 at 20.5%. How to brief Motion like this, step by step, is in the article below ↓ polymarket.com/event/next-cl…