$12,760 in profit on Polymarket from the one Claude workflow everyone skips because it sounds boring. > $11,045 last month > $3,233 week > $880 today the workflow: alerts only, no decisions. his profile: his bot pulls weather api and the Binance feed on a loop. it scores each Polymarket weather and BTC window against actual forecast data and live price action. when the implied probability drifts more than 10x from what the data says, it pings him. he reads the alert, opens the market, hits enter. claude never picks. claude never sizes. claude never decides when to exit. it watches two specific feeds and shouts when the gap opens. one extension nobody is talking about: stacking his alerts into a parlay. a single weather alert returning +400% is one trade. a weather alert and a BTC alert stacked into one parlay position is one trade where each correct leg multiplies the next. same two signals, different math. bot that does the stacking: add his wallet inside, every alert he hits shows up in your stack instantly. scanner in. parlay out.

$12,760 in profit on Polymarket from the one Claude workflow everyone skips because it sounds boring. > $11,045 last month > $3,233 week > $880 today the workflow: alerts only, no decisions. his profile: his bot pulls weather api and the Binance feed on a loop. it scores each Polymarket weather and BTC window against actual forecast data and live price action. when the implied probability drifts more than 10x from what the data says, it pings him. he reads the alert, opens the market, hits enter. claude never picks. claude never sizes. claude never decides when to exit. it watches two specific feeds and shouts when the gap opens. one extension nobody is talking about: stacking his alerts into a parlay. a single weather alert returning +400% is one trade. a weather alert and a BTC alert stacked into one parlay position is one trade where each correct leg multiplies the next. same two signals, different math. bot that does the stacking: add his wallet inside, every alert he hits shows up in your stack instantly. scanner in. parlay out.

there's a reason the biggest Claude deployment in the world calls it an advisor while you're still treating yours as a trader. Mario Rodriguez ships Claude to millions of GitHub developers. Brad Abrams builds the platform under it. neither of them lets the model make the final call. their core architectural commitment is what they call the "Advisor strategy": Claude reviews, suggests, flags. a human or a deterministic rule decides. every layer of GitHub's stack enforces this: > offline evals filter before launch. > a harness wraps the model. > caching keeps cost predictable. all of it exists to make sure Claude advises and something deterministic decides. i landed on the same shape independently, building a Polymarket bot. the first version asked Claude to predict prices and trade them. it lost money. the second version stripped Claude down to one job: review the thesis i had already written and find the strongest reason to reject it. 90 days of running the new prompt. 28 entries blocked. 44 would have lost me $5,360 if i'd taken them. the refused trades made more money than the executed ones. GitHub does this for millions of developers. you should be doing it for one Claude call. build the harness around your prompt before you build anything around your bot. advisor in. trader out.

there's a reason the biggest Claude deployment in the world calls it an advisor while you're still treating yours as a trader. Mario Rodriguez ships Claude to millions of GitHub developers. Brad Abrams builds the platform under it. neither of them lets the model make the final call. their core architectural commitment is what they call the "Advisor strategy": Claude reviews, suggests, flags. a human or a deterministic rule decides. every layer of GitHub's stack enforces this: > offline evals filter before launch. > a harness wraps the model. > caching keeps cost predictable. all of it exists to make sure Claude advises and something deterministic decides. i landed on the same shape independently, building a Polymarket bot. the first version asked Claude to predict prices and trade them. it lost money. the second version stripped Claude down to one job: review the thesis i had already written and find the strongest reason to reject it. 90 days of running the new prompt. 28 entries blocked. 44 would have lost me $5,360 if i'd taken them. the refused trades made more money than the executed ones. GitHub does this for millions of developers. you should be doing it for one Claude call. build the harness around your prompt before you build anything around your bot. advisor in. trader out.

Anthropic's CPO just killed the boundaries between three Claude products. you should kill them inside your prompt too. Mike Krieger dropped a line most builders skipped: > "we're shipping our harness strategy, not a product." he doubled down later: the separation between the three products is "a broken abstraction we need to fix." Anthropic spent a year separating Claude AI, Cowork, and Claude Code into distinct surfaces. their own CPO now says the separation was the bug. they're collapsing it into one harness. the same abstraction is broken inside every Claude prompt you've written. you ask Claude to predict, pick, size, exit, and audit: five jobs glued together pretending to be one workflow. then you wonder why nothing comes out clean. i ran into this on a Polymarket bot. the prediction-layer version asked Claude to do everything. it lost money. then i stripped the prompt down to one job: veto an entry i had already decided to take. nothing else. no picking, no sizing, no exit logic. over 90 days the new prompt blocked 128 trades: 44 of them would have lost me $5,360. the bot's blocked trades saved more dollars than its approved trades earned. Krieger's lesson at the product layer is the same lesson at the prompt layer. - one harness. - one job. - fewer broken abstractions stitched together. Anthropic is collapsing their stack. you can collapse yours tonight. stop stitching five jobs into one Claude call.

the bottleneck Karpathy keeps pointing at isn't the model. it's the file you wrote three months ago telling it how to work. his whole framing: code isn't the verb anymore. > you express your will to agents 16 hours a day. > running more of them in parallel is the unlock. > and he drops this line halfway through, no fanfare: "now you can have optimization over the instructions." that's the sentence. the instructions are the new bottleneck. and almost nobody has a process for optimizing them. i built one for myself. it reads 100 of my past Claude Code sessions and rewrites my CLAUDE.md from what the transcripts show, not from what i thought i wrote. one Saturday morning later: a 38-line memory file. 73% of my old CLAUDE.md got deleted. most of it was one-off corrections that hardened into permanent rules months ago and were quietly steering every agent i ran. four behavior patterns surfaced that i had never written down. the loudest one: my prose corrections outnumber code corrections 8.2 to 1. that ratio was running silently through every session for 90 days. you don't need a smarter model. you need an instruction file that still describes how you actually work. listen to the podcast for the framing. open the article for the script that does the optimization.

every month you use Claude Code, your CLAUDE.md gets longer and worse at the same time. - one-off corrections harden into permanent rules. - context ages out. - you contradict a line you wrote and forgot. - nobody audits any of it. Dreaming is the audit: it reads up to 100 past sessions and rewrites the memory file from what you actually did, not what you once typed. here's what nobody says out loud: > the raw material it needs already sits on your disk as JSONL > the official version is gated behind Managed Agents, enterprise-priced > the mechanism itself is 80 lines of Python and a 12-line rubric i didn't request access. i rebuilt it over a Saturday morning and ran it on 90 days of my own sessions. one Opus call and out came a 38-line memory file. 73% of my old CLAUDE.md deleted. four behavior patterns surfaced that i had never written down, including one where i correct Claude's prose 8.2x more than its code. watch the keynote for what shipped. read the article for the version you can run tonight. your CLAUDE.md isn't too short. it's too old.

you can write a perfect CLAUDE.md today and it will still be lying to Claude in three months. not because you wrote it wrong. because the file is static and your work isn't. a correction you made once becomes a permanent rule. a project name goes stale. a preference you typed in January quietly contradicts one you typed in March. the file never changes. you do. writing a good CLAUDE.md is a solved problem. keeping it true to how you actually work is the part nobody has a process for. i gave myself one: it reads 100 of my past Claude Code sessions and rewrites the file from what the transcripts show, not from what i remember typing. the result on my own setup: > 73% of my CLAUDE.md got deleted on the first pass > most of it was one-off corrections that had hardened into rules > four patterns i had never written down got surfaced instead > one of them: i correct Claude's prose 8.2x more than its code the keep list was 38 lines. the delete list taught me more. learn to write the file from the video. then read the article for the part that comes after: keeping it honest. a CLAUDE.md you never audit isn't a memory file. it's a museum.

you can write a perfect CLAUDE.md today and it will still be lying to Claude in three months. not because you wrote it wrong. because the file is static and your work isn't. a correction you made once becomes a permanent rule. a project name goes stale. a preference you typed in January quietly contradicts one you typed in March. the file never changes. you do. writing a good CLAUDE.md is a solved problem. keeping it true to how you actually work is the part nobody has a process for. i gave myself one: it reads 100 of my past Claude Code sessions and rewrites the file from what the transcripts show, not from what i remember typing. the result on my own setup: > 73% of my CLAUDE.md got deleted on the first pass > most of it was one-off corrections that had hardened into rules > four patterns i had never written down got surfaced instead > one of them: i correct Claude's prose 8.2x more than its code the keep list was 38 lines. the delete list taught me more. learn to write the file from the video. then read the article for the part that comes after: keeping it honest. a CLAUDE.md you never audit isn't a memory file. it's a museum.

Anthropic just shipped Dreaming agents review their own sessions between runs, surface recurring mistakes, and rewrite their memory so the next session doesn't repeat them. Karpathy wrote CLAUDE.md for the same reason. > silent wrong assumptions > over-complication > orthogonal damage. his 4 rules cut Claude's mistakes from 41% to 11%. it was memory between sessions, written in 4 lines of markdown. Anthropic's platform team walked the architecture on stage last week at Code with Claude SF. Dreaming runs the same loop, automated, at platform level. CLAUDE.md is memory you write. Dreaming is memory the agent writes about itself. i ran Karpathy's 4 rules across 30 codebases for 6 weeks and added 8 more. it cut my mistakes another 8 points. all 12 are in the article. the 12 rules don't compete with Dreaming. they sit on top of it.

Anthropic just shipped Dreaming agents review their own sessions between runs, surface recurring mistakes, and rewrite their memory so the next session doesn't repeat them. Karpathy wrote CLAUDE.md for the same reason. > silent wrong assumptions > over-complication > orthogonal damage. his 4 rules cut Claude's mistakes from 41% to 11%. it was memory between sessions, written in 4 lines of markdown. Anthropic's platform team walked the architecture on stage last week at Code with Claude SF. Dreaming runs the same loop, automated, at platform level. CLAUDE.md is memory you write. Dreaming is memory the agent writes about itself. i ran Karpathy's 4 rules across 30 codebases for 6 weeks and added 8 more. it cut my mistakes another 8 points. all 12 are in the article. the 12 rules don't compete with Dreaming. they sit on top of it.

a meteorologist got replaced by an AI weather model after 8 years on the job. 6 months later his bot printed $88,018 on Polymarket. $13,800 every month. curve never dipped. his boss still pays the AI subscription xd the trade itself isn't complicated. ECMWF and NOAA push raw forecast data every few hours. Polymarket temperature markets price on consensus narrative, not on the latest model run. the gap is 10x bigger than the book pricing implies. his best: - $5.50 -> $764.19 (+13,797%), NYC 44-45°F - $11.88 -> $1,496.00 (+12,497%), NYC 78-79°F - $18.32 -> $1,794.44 (+9,695%, NYC 63-64°F bot watches both APIs 24/7 and enters the second the spread opens. profile: this is Karpathy's CLAUDE.md thesis on a different table. > silent wrong assumption: "the market price already contains the forecast." it doesn't. > over-complication: stacking 7 indicators when one API gap is the whole edge. > orthogonal damage: hedging into correlated markets that resolve on the same event. he stripped all three. single markets, Kelly-sized, exit on resolution. that's the entire system. the leverage layer on top: same 4-5 entries combined into one parlay. four 95% legs multiply into a payout that doesn't fit a single position. it's not edge, it's compounding on edge that already works. the API gap is the alpha. the parlay is the multiplier. he used it: operator runs the API and knows why he's in every position. tourist runs the bot link without the API and finds out at resolution. the 12 rules I added on top of Karpathy's 4 are in the article. weather doesn't lie. the book does.

a meteorologist got replaced by an AI weather model after 8 years on the job. 6 months later his bot printed $88,018 on Polymarket. $13,800 every month. curve never dipped. his boss still pays the AI subscription xd the trade itself isn't complicated. ECMWF and NOAA push raw forecast data every few hours. Polymarket temperature markets price on consensus narrative, not on the latest model run. the gap is 10x bigger than the book pricing implies. his best: - $5.50 -> $764.19 (+13,797%), NYC 44-45°F - $11.88 -> $1,496.00 (+12,497%), NYC 78-79°F - $18.32 -> $1,794.44 (+9,695%, NYC 63-64°F bot watches both APIs 24/7 and enters the second the spread opens. profile: this is Karpathy's CLAUDE.md thesis on a different table. > silent wrong assumption: "the market price already contains the forecast." it doesn't. > over-complication: stacking 7 indicators when one API gap is the whole edge. > orthogonal damage: hedging into correlated markets that resolve on the same event. he stripped all three. single markets, Kelly-sized, exit on resolution. that's the entire system. the leverage layer on top: same 4-5 entries combined into one parlay. four 95% legs multiply into a payout that doesn't fit a single position. it's not edge, it's compounding on edge that already works. the API gap is the alpha. the parlay is the multiplier. he used it: operator runs the API and knows why he's in every position. tourist runs the bot link without the API and finds out at resolution. the 12 rules I added on top of Karpathy's 4 are in the article. weather doesn't lie. the book does.

head of Claude Code doesn't write code anymore. he reviews it. he said this on CNBC this week. if your plugins are doing the same job to your tokens before Claude even starts, you're not reviewing. you're paying for a fight. i put an HTTP proxy on 14 Claude Code setups for 11 days: one Edit fired 7 PostToolUse events. 34% of every token bill burned on plugins trampling each other. there's a 3-line fix in the article.

Boris Cherny ships 30 PRs a day from his phone. his actual words: he hasn't typed code by hand in 2026. the title software engineer is going away. it's just "builder" now. most engineers heard "we're obsolete." it's not what he said. his framing: > coding is solved when your agent stack actually works > the bottleneck moved from typing to orchestration > the winners are the ones who can wrangle agents > the losers will blame the model builders need infrastructure. that is what nobody is telling you. I spent 11 days putting an HTTP proxy on 14 Claude Code setups. 34% of every token bill burned on plugins fighting each other before Claude generated a single useful output. one setup ran 7 PostToolUse fires on a single Edit. 7x the tokens. developer thought the model was "running slow." it wasn't. the gap between you and Boris isn't talent. it isn't tokens. it's the 30 lines in his settings.json that yours doesn't have. 3-line lock-file fix in the article. open it. coding may be solved. your settings.json is not.

Karpathy threw a grenade at every senior engineer who still treats LLMs as a toy. his actual words: the worst thing an expert can do right now is reject them. most experts read it as a threat, but it's advice. his framing: > the gap between "AI tools are bad" and "AI tools are useful when used right" is professional discipline, not capability > agents have cognitive deficits. they fail in ways nothing in the training set anticipated > the experts who reject LLMs lose to experts who learn to wrangle them > "models have so many cognitive deficits. but you can route around them" routing around the deficits is what CLAUDE.md was invented for. Karpathy himself wrote 4 rules. across 30 codebases they took my Claude error rate from 41% down to 11%. solid drop. but his rules pre-date the slop era going public. I bolted on 8 more, tuned to the failure modes that surfaced after January. got it down to 3%. a CLAUDE.md does not raise Claude's IQ. it lowers his slop floor. that is the entire game. open the article underneath. the model is not the bottleneck. your config is.

the man who coined "vibe coding" just said he's never felt more behind as a programmer. Andrej Karpathy went from writing 80% of his code to delegating 80% to agents. In one month. his core thesis: vibe coding raised the floor, agentic engineering raises the ceiling. what shifted in his workflow: > code is not the verb anymore. "I express my will to my agents 16 hours a day" > idle tokens = you're the bottleneck. run more agents in parallel > jagged ghosts, not animals. spike on what RL rewarded, fail on the rest > you can outsource thinking. you cannot outsource understanding his 4 CLAUDE.md rules from January are the floor for this transition. they cut my Claude mistakes from 41% to 11% across 30 codebases. but they were built for January's problem. the May 2026 agent-orchestration problems didn't exist yet. so I added 8 more rules. got the mistake rate to 3%. watch the lecture, read the breakdown. stop blaming the model. fix your claude.

this trader hit 370/1 wl and maded $8,931 in 2 weeks. > $1,247 today > $4,318 past week > 99.7% wr. everyone trades BTC 5-min. he found a quieter edge. china weather markets. guangzhou. beijing. wuhan. that's it. his stack: > CMA api (china meteorological administration) > ECMWF for cross-check > claude reads both, compares to polymarket price > enters NO at 90-96c on outcomes already 99% confirmed > Kelly criterion sizes the bet > waits for resolution f* = (p − m) / (1 − m) best trades: - guangzhou 22°C+ no, 94c -> $1.00 ($3,200 -> $3,403) - beijing 8°C+ no, 96c -> $1.00 ($4,150 -> $4,323) - wuhan 15°C+ no, 91c -> $1.00 ($1,890 -> $2,077) each trade pays 4-10%. he runs 25-30 a day. compounds daily. curve hasn't dipped once since april. straight line up. his profile: here's the asymmetry copy traders miss: - 4 trades at 96c copied separately = +4% x 4 = +16% - same 4 stacked into one parlay = +100-300% in one position stop copying his trades one by one. stack them. do it via most consistent weather wallet on polymarket. nobody's watching him yet. don't wait until the spread closes.