anthropic and openai put the real manuals for claude and chatgpt on github, for free most people who use them every day have never opened one. 10 repos. 783k stars between them. prompting courses, agent code, the coding agents themselves, and an openai model that runs on a 16gb laptop. here's what's inside: learn -> how the labs prompt their own models prompt-eng-interactive-tutorial · anthropic's 9-chapter prompting course, with exercises and an answer key github.com/anthropics/prompt… claude-cookbooks · copy-paste recipes: rag, sub-agents, pdfs, vision, prompt caching github.com/anthropics/claude… openai-cookbook · openai's own examples and guides for its api github.com/openai/openai-coo… run -> the agents themselves claude-code · claude in your terminal. reads your codebase and handles the routine work github.com/anthropics/claude… codex · openai's coding agent, runs locally on your computer github.com/openai/codex gpt-oss · openai's open-weight models. the 20b one runs in 16gb of memory github.com/openai/gpt-oss build -> your own agents openai-agents-python · agents, handoffs, guardrails, voice. works with 100+ models, not just openai's github.com/openai/openai-age… shared -> the plumbing both labs use skills · the folders claude loads to make word, excel, powerpoint and pdf files github.com/anthropics/skills mcp servers · the "usb-c port" that plugs ai into your files, apps and tools github.com/modelcontextproto… agents.md · one file that tells any coding agent how your project works github.com/openai/agents.md and that's the funny part: these two fight over everything. except the plumbing. skills started at anthropic. now chatgpt, codex, cursor, gemini cli and copilot load them. mcp started at anthropic. now it lives at the linux foundation, in a home openai co-founded. agents.md came from openai. 60k+ projects use it. and the skills repo has 177.5k stars. claude code itself has 148.8k. the manual got bigger than the product. stop saving prompt screenshots. read what the labs wrote. save this for the weekend.

WTF, I JUST REPLACED A $500K DEV AGENCY WITH A $200/MO GROK BOT SETUP Here is what my studio looks like in the cloud. Chief of Staff takes the brief and routes the work: > ARCHITECT designs the database and API from my prompts > FRONTEND (UI) builds the components and pixel-perfect screens > FRONTEND (STATE) wires the logic and connects the endpoints > BACKEND (CORE) writes the business logic and handles routing > BACKEND (DATA) manages the DB queries and server config > QA breaks the code, finds the bugs, and forces a rewrite BEFORE I ever see it 6 agents -> 1 shared operation -> automatic handoffs -> human approval The secret isnt that these are six smart bots. It is that they share one persistent cloud environment The backend bots dont need to make pull requests -> the frontend bots already see the code in the shared folder No sync calls with a PM. No waiting on dev ops. Each bot keeps its own workflow, passes finished work forward, and I just approve the final build Launching a product used to mean raising a seed round, signing a lease, and hiring a team. Mine took an evening

a 20-year-old put $1,000 on Trump saying two specific words on live TV. it paid $95,000 one question: will he say "fake news" this week? here's the part most people miss: he wasn't gambling. he was pricing a habit "Trump says fake news" sounds random if you've never listened to him speak. if you have, it's closer to betting the sun comes up. the market priced a personality trait like it was a coin flip that's the whole edge -> the crowd prices things by how silly they sound, not by how likely they are the clip is the best part. he buys it, then can't sit still. pacing outside, checking the stream, staring at the podium. Trump opens his mouth, says it, and the kid comes off the couch and that exact market runs every single week the one for August 3-9 is live right now and closes Saturday: -> "no no no" -> 95%, the crowd treats it as already done -> "moon" -> 29c -> "six seven" -> 18c -> "stupid question" -> 14c -> "epic fury" -> 13c -> "karoline" -> 13c the cheap ones at the bottom are where that edge lives. sometimes the crowd is right and they're genuinely unlikely. sometimes it just prices them low because they sound stupid, and that gap is the entire game live board here: polymarket.com?via=banff save this. the biggest edges hide inside the trades everyone else laughs at

Creator of Claude Code: "Onboarding used to take two or three weeks. It's now about two or three days" That's what happened inside Anthropic when new engineers stopped asking humans questions and started asking the AI instead. The whole talk is 24 minutes of tips from the guy who built the thing, and almost none of it is what people repeat online: -> don't ask it to build the feature. ask it to make a plan first and wait for your approval. that one line fixes most bad results -> give it a way to check its own work -> a test, a screenshot, anything. then it iterates on its own and lands close to perfect instead of stopping at "fine" -> he doesn't type his prompts. he dictates them out loud, like talking to another engineer -> nobody programmed it to read git history. it just knows. he says the same about half the things it does save this. 24 minutes from the person who built it beats a hundred threads from people who didn't

a piano holds 18 tons of string tension. guitar strings snap at 90 kilos he put guitar strings on a piano anyway the numbers said it couldn't work, and he listed the reasons before he started: 210 strings to buy. a professional tuner charges around $20 per string, so he was about to do $4,000 of labour by hand, having barely ever restrung his own guitar the strings are too short for a piano, so he had to tie them together end to end and not one guitar tuning peg fits a piano. it took him days of searching to find something that would then he started threading. thinnest guitar string in the spot where the thinnest piano string sits, next one after that, no measurements, no math, just feel he ran out of strings partway through, because tying them together ate more length than he calculated but the tied ones rang clean. that was the moment it stopped being a joke then he tuned all of it the only way he knows how -> a phone app and a guitar pick, string by string, on an instrument built to hold the weight of a small truck and it played Wonderwall that's the whole format. no plot, no twist, just a number that says impossible and a guy who does it anyway save this. a piano that should have torn itself apart played the most overplayed song on earth unrelated, but worth a read if you're into prediction markets, fresh piece from Boba: