vivi

I GAVE 3 CLAUDE FABLE 5.2 AGENTS ONE NIGHT AND $150 IN TOKENS TO BUILD ME A TRADING TERMINAL FROM AN EMPTY FOLDER. I went to bed at 11 PM. At 6:52 AM there was a new app in my Dock, 212 of 212 tests green, a token bill for $146.30, and the $300 I left in it had turned into $1,912 The folder had one file in it. rules.md, three lines: "the agent that writes the code never checks it" "nothing touches real money until every test passes" "never more than 10% of the balance on one trade" Who did what: Architect wrote the plan at 11:04 PM. 9 panels, 4 price feeds, one order button. After that it only said yes or no. Builder wrote 63 files and rewrote the order book panel 4 times. Breaker had one job: break everything. It found 38 bugs. At 2:17 AM it fed the terminal a fake 38% price spike and the stop-loss didn't fire. Builder fixed it by 2:28. At 4:06 AM the last test went green and Architect switched it from paper to my real $300. 19 trades by morning, 15 green. Best one: $GOAT, +$412.40. 7 hr 52 min. $146.30 in tokens. $300 to $1,912.

THIS IS THE CHEAPEST SECOND BRAIN YOU WILL EVER BUILD no Obsidian, no Notion, no database. a handful of .md files on your disk and one index the agent reads before you type anything. the index is CLAUDE.md in the root. Claude Code loads it into the system prompt at session start, so the map is already in its head. ~/brain $ mkdir notes people projects && touch CLAUDE.md here's the split: ▸ notes/ holds one file per topic, named for the topic ▸ people/ holds one card per person, who they are and what you owe them ▸ projects/ holds status, where you stopped, the next step ▸ MEMORY.md, LEARNINGS.md and decisions.md sit in the root as the core, the lessons and the audit log the rule that makes it work: one file, one topic. tax-policy.md, not notes.md. ivan-petrov.md, not note-12.md. lowercase and hyphens, always. the part that makes it fast: nothing gets installed. these are text files. any editor, any agent, any machine, and they outlive whatever app you would have picked instead. you ask "what did we decide on taxes?" and it goes map, then notes/tax-policy.md, then an answer with the file cited. ask about something you never wrote down and it tells you it is not in there. that second one is the whole point. a clean structure is what lets it say "missing" instead of inventing a number. you can pin a file with @MEMORY.md inside the map and it stays in focus every session. two files at most, or you are back to dumping everything into the prompt. one honest warning: the map is the product, the folders are just storage. a sloppy CLAUDE.md gets you an agent that guesses, and it guesses confidently. five minutes of folders tonight, filled as you go. no migration weekend. save it, this is the cheapest memory your agent will ever have.

a rebuilt function matched the original byte for byte and was still wrong the comparison tool lined up the instructions and never looked at the number they pointed to. the original checked player power against 128.0. the rebuild checked against 0.0. every byte matched, the meaning did not the audit after it went through 1,548 floating-point references and found twelve more wrong ones, in five functions that had already been accepted that is the part the install guides skip. the model reads, explains and rewrites. the check decides whether any of it is true, and the people doing this seriously have a name for the check. they call it the oracle. doesn't matter if you're rebuilding a feature, porting a codebase or just letting an agent grade itself, the oracle is the whole job the three kinds, and how to break them on purpose: > byte oracle - compile your version, compare machine code. DX-Ball's audio balance helper: 3,205 test cases pass and all 63 compiled bytes reproduced. that is proven, not "looks right" > behaviour oracle - same inputs into both, compare outputs. why Windows Calculator answers 220 to 200 + 10%, cross-checked against Microsoft's own published source with the exact version written down > runtime oracle - build the whole thing and use it. the Chrome dino speed rule starts at 6, adds 0.001 per update, caps at 13, and they ran it: 10.0 after 4,000 updates, 13.0 after 10,000 > plant 3 failures before you trust a green run. one wrong constant, one swapped branch, one off-by-one. if any of them passes, the oracle is broken and every green result before it is suspect > room A reads with REA and writes SPEC.md - behaviour, inputs, outputs, edge cases, unknowns. no code, no addresses, no names from the decompile > room B is a fresh session in a separate folder that gets SPEC.md and nothing else. it never sees the analysis, which is the entire point of a clean room > hold the builder to a measurable end: every case in the spec passes, all 3 planted failures rejected, nothing outside the folder touched, stop after 25 turns the Adobe story everyone repeated is the proof. nobody decompiled anything. ArtCraft shipped twelve apps written from scratch in Rust, and FilmCraft's contributor rules forbid reading, disassembling or copying anything inside an Adobe bundle at all. you observe the behaviour by using the app. the people who rebuilt Adobe's workflow banned themselves from reading Adobe. study the original, ship your own. an output is not a right, and reading a program gives you no claim on its code, its data or its name. the board in my clip is a visual concept for software you own or may examine. the full guide, the two-room prompts and the planted-failure pack are below.

ask claude whether a trading rule works and you get an answer in 4 seconds it has never seen the chart. it has never run the test. it is answering from memory, and a smarter model with no data is just a more convincing guess the fix is not a better model. it is taking the grading away from the thing that had the idea one chat window picks the idea, runs the test, grades its own work and forgets all of it when you close the tab. that is four jobs and zero separation. a desk splits them. doesn't matter if you're testing strategies, research claims or anything you can be wrong about expensively, this shape is the point the 5 roles: > lead on Opus 5.5 turns the question into a plan and decides what runs next. it never grades its own output > scout on Sonnet 5.5 checks the data before a single test runs - delisted companies included since 1998, fundamentals stamped by filing date instead of quarter end. survivorship and look-ahead bias are the two quiet ways a backtest lies to you > sweeper on Sonnet 5.5 runs 20 configs from one prompt and returns every cell, failures included, so you can see whether the wins form a plateau or one lucky number > skeptic on Opus 5.5 exists only to break what survived - held-out years, nearby settings, the worst drawdown year. the strongest model goes here on purpose, because generating ideas is cheap and killing bad ones is not > librarian on Haiku 5.5 pulls from 175 documented strategies instead of from a feeling about what should work > research.md logs every verdict before any summary, and NO EDGE gets logged and never retested without a new reason > the split is by time, at 2015, never random. shuffling time leaks the future into the past the honest half, because this is money and not a side project: the 200-day numbers in that course are the author's published run, gross of costs, so fees and slippage can erase a thin edge. backtests do not model taxes or slippage, the agents never place an order, and a beautiful equity curve marked NO EDGE is still NO EDGE. paper first, always. the desk in my clip is a visual concept, not a live account. the course below has the agent files, the four team rules and the full setup.

🚨 The software industry isn't ready for what's coming. Imagine taking almost ANY software, understanding how it works, and building your own version with AI. Sounds impossible? Not anymore. Meet REA (Reverse Engineer Anything) — one of the wildest open-source AI tools right now. And when combined with Claude Code, things get absolutely insane. Here's why everyone should be paying attention 🧵👇 01 / AI CAN NOW INVESTIGATE SOFTWARE Previously, reverse engineering required serious technical knowledge, expensive tools, and countless hours. Now AI agents can use REA to: → Analyze software architecture → Trace hidden functionality → Examine binaries and dependencies → Understand how complex features work → Collect evidence behind their findings What used to be an extremely technical process is becoming accessible through natural language. 02 / FROM ANALYSIS TO REBUILDING Here's where it gets interesting. REA helps your AI agent understand how software behaves. Claude Code can then turn those findings into an independent implementation. The workflow is surprisingly simple: 🔍 Investigate the software 🧠 Extract its core functionality 📝 Generate a detailed specification ⚡ Build your own version ✅ Test, verify, and improve All within an AI-powered development workflow. 03 / THE BIGGER PICTURE Think about what this could mean. → Building alternatives to expensive software → Understanding complicated applications → Creating custom tools for personal workflows → Accelerating development dramatically → Making software research more accessible And we're still in the early stages. Of course, reverse engineering doesn't remove copyright, licensing, or other legal restrictions. But here's the part most people are missing: The real breakthrough isn't copying software. It's giving AI the ability to understand existing systems and use that knowledge to build something new. We're moving toward a world where understanding complex software no longer requires an entire team of specialists. The barrier between having an idea and building a working product is getting smaller every day. And tools like REA are pushing that transformation forward. Closed-source software isn't disappearing overnight. But the way we analyze, understand, and recreate functionality is changing forever. REA + Claude Code is a combination worth watching. The next generation of software development is going to look VERY different. **Are we witnessing the beginnin

you don't need a bigger plan. 87,500 of the 100,000 tokens a task burns go to work your bot already did yesterday's chat dragged into today, full pages read for one price, polling a source that never changed, and research rebuilt from scratch because nothing saved where it came from seven fixes, and the one that matters most is the one that costs more, not less doesn't matter if you're running research, monitoring or reporting on a schedule, audit before you pay for more capacity the 7 fixes, with where the tokens go: > 01 start from a file, not yesterday's thread. goal, decisions and next action in HANDOFF.md, under 400 words, then open a clean conversation. 28k to 1.5k > 02 read what the task needs. pull the relevant passages first, keep the full original in sources/, open more only when missing context would change the answer. 22k to 3k > 03 stop paying for "nothing changed". compare against the last snapshot before any analysis, and never let a failed check count as unchanged. 18k to 1k > 04 give finished work an expiry rule. store checked_at and a refresh rule, reuse what is still valid, and keep the original verification date instead of relabelling old facts as new. 12k to 1k > 05 one owner per assignment. one coordinator, separate output files, no source collected twice without a recorded reason. 10k to 0.5k > 06 cap retries at 2. save progress at real boundaries, log the exact error, and never restart the four briefs that already passed because the fifth page would not load. 7k to 1.5k > 07 make "done" mean usable. checklist before drafting, PASS FAIL UNKNOWN against evidence, one targeted fix pass, NEEDS REVIEW if a check still fails. 3k to 4k read that last line again. verification is the only row that goes up, and it is the reason the other six are worth anything. a bot saying "verified" is not evidence. the 10x is arithmetic, not a benchmark: 8x more assignments processed times 1.25x acceptance equals 70 accepted results instead of 7, on the same million tokens. the author says plainly that real usage stays unknown until a meter measures it, and the board in my clip is his table, not my run. the full course, the copy-paste prompts and the 4 github tools are in the article below.

you do not prompt an employee. you brief one and the brief that decides whether any of this works is 4 lines long, pasted once, before the first job most people skip it and then wonder why the agent writes in a voice that isn't theirs, chases the wrong customers, and one day sends something it should have shown them first. that is not a model problem. nobody told it who it works for. doesn't matter if you're handing off research, outreach or reporting, start here the 4 lines: > Me: your name and what you actually do, in one sentence a stranger would understand. not your title > My customers: who buys from you and who doesn't. "dental clinics, independent, not chains" beats "small businesses" > My voice: two lines you actually wrote. not an adjective. it copies rhythm from examples and ignores "professional but friendly" > Never: contact anyone, post, or pay without my yes. this is the one that keeps the other three from becoming expensive then the jobs, one a week: > give each one a day and a time instead of remembering to ask. "every monday at 9:00, 25 new leads" and "every friday at 16:00, the week on one page" > demand the artifact, not the answer. a spreadsheet with an empty approval column, a one-page pdf, a report ending in three lines of next week > make it show receipts for every claim. in the clip: 240 read, 192 dropped, 48 kept, 12 claims each linked to a source and 0 invented the graph in the clip is my own concept and every number on it is simulated, the backtest included. nothing on that screen is a result i'm claiming. the guide below is a paid partnership built around one product, so the timings in it are the author's runs. the four-line brief and the one-job-a-week order cost nothing and work with whatever agent you already pay for.