weekly limits reset at 3am. my agents were already waiting by the time i woke up: 20% of the session gone, the overnight report written, two decisions sitting in the queue for approval i read 11 lines over coffee and said yes twice. that was the entire standup the terminal is the new office and it clocks in before you do

a reply is worth 27 likes. your reply to a comment is worth 150. these numbers sit in X's public code and almost nobody has ever opened the file 5 days ago my account was a graveyard: 46K followers, 600 views a post. instead of quitting i read the ranking code on github and rebuilt the whole account around the actual math [ what the code says ] > a like weighs 0.5. a reply weighs 13.5. one comment = 27 likes > the author answering a comment and getting a reaction: 75. that's 150 likes for typing one reply > a mute or "show less often": -74. one annoyed reader erases everything two viral replies built > the feed sorts every account into ONE community. blurry accounts get recommended to nobody [ what i changed ] > volume: 2 posts a week → 5-8 a day. every post is a lottery ticket, the algorithm counts signals, not effort > one cluster: dropped crypto content completely. picked a lane > the format mix: it used to be one article a week quoted with clips, that was the entire account. now it's news quotes, repo drops, personal screenshots, stories > replies: i answer every single comment. the math above is the reason [ the honest part ] 5-8 posts a day is not writable by hand while also running a company. so i eat my own cooking: i fed @voxlyink my whole post history, winners and flops, and it drafts in my voice from what actually performed. i cut, sharpen the hook, hit post. volume is a system, not willpower [ 5 days later ] > views up 3x, curve still bending > likes from zero to double digits > best performer of the week: a screenshot of my usage limits. personal receipts beat essays, the code explains why not a success story yet. a lab notebook. next update when a post breaks 10K the algorithm isn't a mystery. it's a spec sheet, it's public, and your competition isn't reading it either

a YC partner just quit handing out money to go work on Anthropic's compute team and everyone read it as a career update read the room. in one week: > Anthropic reportedly starts chip talks with Samsung > reports of an S-1 being prepped for an IPO as early as October > free Fable extended a third time while usage graphs go vertical > and now YC partners are leaving the easiest job in tech to work on "recursive self-improvement" when the people whose whole career is spotting winners stop betting on startups and join the lab itself, that's not a hire. that's insider information walking through the front door the quiet company stopped being quiet

mars doesn't need a robot that folds laundry 1X shipped hands for the kitchen this week and everyone clapped. meanwhile the actual frontier has zero robots designed for it by anyone outside three government labs so we're going to fix that. MARS ROBOT MARATHON. details below [ the setup ] crewed Mars flights are closer than most people think. but humans land last. first go the scouts, the drillers, the cave divers, and that fleet does not exist yet > Perseverance cost $2.7B and took thousands of engineers a decade > Ingenuity was supposed to fly 5 times. it flew 72. the assumptions are always wrong in the builder's favor > AI already designed a working rocket engine that got printed and test-fired > designing a Mars scout is now a problem one person with an agent team attacks from a bedroom the tools got democratized. the ideas didn't catch up yet. that's the opening [ the spec, non-negotiable ] > walks on terrain that eats wheels for breakfast > climbs crater walls and 40 degree slopes > drills, because everything interesting on Mars is underground > dodges obstacles alone. instructions from Earth take up to 20 minutes one way, a rock doesn't wait > descends into lava tube caves and maps them [ why caves are the whole game ] > Mars gravity is a third of Earth's, so its lava tubes grow monstrous. wide enough to hide a city block > they block most of the surface radiation > temperature inside stays stable while the surface swings 100 degrees a day > if life still exists on Mars, it's not sunbathing on the surface. it's down there in the dark whoever maps the caves first owns the map everyone else lands by [ the marathon ] design the scout that goes first. experiment with shapes: spider on a tether, snake that swallows tunnels, a swarm with one nuclear mother and ten expendable kids. mars doesn't care about earth aesthetics > drop your design in the replies: sketch, render, napkin drawing, AI blueprint, anything > best 3 get a full breakdown in a dedicated post > the winner gets turned into a complete engineering blueprint, my agent team does the drafting i'm dropping my own MK-I first your move, engineers

every AI lab is losing money on you right now, on purpose, and the receipts are public [ the receipts ] > a dev posted his dashboard: $200 subscription, $8,545.93 in actual usage. 11B tokens in 30 days. a 97% discount, and he's not an outlier > a team lead posted their first week running both Sol and Fable: 30% of the entire bill was one model > Anthropic extended free Fable 5 access three times in a row and raised Claude Code limits 50%. through July 19, again > xAI ships Grok 4.5 at roughly half of Opus pricing and Musk moves Tesla and SpaceX onto it, citing token costs in the memo > OpenAI answers with GPT-5.6 the same week, selling "performance per dollar" right in the announcement [ what's actually happening ] this is the uber playbook, token edition. burn investor money to make the product feel free, let your workflow grow roots around it, flip the meter when leaving costs more than paying they're not fighting for this month's revenue. they're fighting over whose model your muscle memory belongs to when the subsidies die [ what to do about it ] > milk the window: the $200 plans are the most mispriced product in tech right now > build portable: skills, briefs, and configs that move between models, so the meter flip never owns you > route by price: expensive model plans and reviews, cheap models grind. the bill drops ~40% before any subsidy the free lunch has a countdown. eat fast, but bring a container full routing playbook is in my pinned

the CEO of a note-taking app just open sourced the AI skills he was quietly using in his own vault kepano runs Obsidian. millions of users, tiny team, the guy who wrote the "File over app" essay turns out for weeks he's been running Claude Code inside his personal vault with 5 private skills: 1) markdown that respects wikilinks, so nothing breaks the graph 2) Bases queries Claude actually writes correctly 3) canvas edits that don't corrupt the file 4) defuddle: feed it any URL, it strips the ads and drops a clean note into your vault 5) each one is a single SKILL.md. drop the folder into .claude/skills/ and Claude just.. speaks Obsidian no plugin store. no account. no cloud. MIT license, shipped under his own name the part i keep thinking about: his essay was called "File over app". this is the sequel, file over agent. your notes stay yours, the AI is a guest that learned your language your second brain just got hands. what it's still missing is a schedule. that part is in my pinned

my Claude works a night shift now. i wake up to finished work and it costs $0 the full night crew, every repo free: [ the foreman ] RuFlow: takes one goal and splits it across parallel agents. research, code, tests, all running at once > [ the foundation ] gstack: YC president's open starter, so agents build on real structure instead of guessing your architecture. my own product started from this repo > [ the free brains ] cc-compatible-models: run Claude Code on DeepSeek, Qwen, Kimi, GLM. configs and pricing for every one, so overnight runs stop costing Opus money > [ the morning report ] ai-memory-vault: templates that make the crew log what it did, what broke, and what's next. you read 10 lines over coffee instead of scrolling terminal history > [ the glue ] no repo, one spec: the goal, what "done" means, when to stop, what to log. paste it at 11pm, read the report at 8am. my exact template is in my pinned [ where everyone fails ] they hand the night shift a vague goal and wake up to confident garbage. the fix is one line: "done means X. prove it with a diff and a test run" the 9 to 5 was never the productive hours. it was just the hours you were awake what did your agents build last night?

Anthropic just extended free Fable 5 for the THIRD time and raised Claude Code limits 50% on top nobody extends a free trial three times out of generosity. you extend it because the usage graphs are too insane to unplug before the paid switch flips first dose free. second dose free. third dose free, with a bigger needle july 19 the meter starts. allegedly

your Claude lies to you every day and you keep thanking it the model that made the work is a terrible judge of the work. it talks itself into "good enough" every single time. that's why your "done" tasks keep breaking the fix is a second agent with zero memory of the effort. copy this: ----- You are the reviewer. You did not write this work. You do not know who did, and you don't care. Here is the spec: [paste the original brief] Grade the work against the spec only. For each requirement: PASS or FAIL with one line of evidence. Any FAIL = the whole submission fails. You gain nothing from being nice. A false PASS costs us money. Do not suggest fixes. Reject and state why. ----- works in Claude Code, Cursor, or a plain chat. the full system with four seats and three ready teams is in my pinned bookmark this. you'll need it the day something "done" breaks in production

you can build the entire "AI team" setup gurus charge $500 courses for with four free repos [ the stack ] 1. ruvnet/ruflo the "runs 60 agents at once" orchestrator. plugs into Claude Code and splits research, coding, and testing across parallel agents. top of GitHub trending for a reason 2. garrytan/gstack YC president's open starter. a clean, opinionated foundation so your agents build on real structure instead of guessing your architecture. my own product started from this exact repo 3. Alorse/cc-compatible-models the cheat sheet for running Claude Code on cheap models: DeepSeek, Qwen, MiniMax, Kimi, GLM. configs and pricing for every single one. this is how your limits stop mattering 4. jaredrhod/ai-memory-vault ready templates for agent memory: your setup stops forgetting what it did yesterday and starts compounding orchestration, foundation, cheap engines, memory. that's a full org chart for $0 which repo did i miss? best replies go into part 2

we spent three years teaching AI to read everything on earth and the most viral product of the week is a font it can't read look at what actually happened: > one guy ships a font readable by humans, unreadable by Fable and GPT-5.6 Sol Ultra > 14M views in a day. not for giving AI a new ability, for taking one away > the same week Claude got a browser to read the web and robots got fingers to touch the world the market just started pricing a brand new asset: spaces where AI can't follow you today it's a font. next it's messengers, exams, contracts, whole platforms. "AI-free" is about to become a paid feature, like "organic" but for information every counter-economy starts as a joke

the app your uncle uses to buy 3 shares of Tesla just flipped Ethereum in DEX volume it took 11 days > Robinhood Chain mainnet is live, already #2# in DEX volume, and today flipped ETH in 24h volume per DefiLlama > TVL crossed $130M, up 17% in a single day > memecoins on the chain passed $244M combined market cap, up 33% today, individual ones printing 700%+ > the launchpad had to PAUSE new memecoin launches because bots were spamming it faster than humans could ape crypto spent a decade building decentralized rails for the masses. the masses walked in through a stock app with a green logo distribution beats ideology. every single time

you think you're watching a beef, you're watching a price war with characters the episode so far: > Musk moves Tesla and SpaceX onto Grok 4.5, then accuses Sam of scamming, right on his own platform > Sam fires back: "the most reliable way to tell 5.6 sol is the best model in the world is that elon is obsessed with me again" > the two biggest egos in AI, arguing in public like a divorced couple at a wedding now the same week, with the drama muted: > xAI ships Grok 4.5 at roughly half of Opus pricing, and the Tesla memo cites token costs, not loyalty > OpenAI answers with GPT-5.6 Sol within days, straight into "best model in the world" claims > Anthropic stays silent, extends free Fable 5 access for the second time, and lets the Opus 5 rumors do the talking.. the quiet one at the table is reloading three labs, one week, and every single move is about price per task, not intelligence they fight for your attention because their models fight for your bill the feud is the marketing. the discount is the product