Arcvale @itsarcvale

IT IS BECOMING F*CKING SCARY A developer just pointed Claude's new Split Agent mode at his repo to see what happens when two models refuse to let each other cut corners He launched a dual terminal session at 1 AM, fed it a 9,000 line codebase overhaul, and watched the live output: * Agent #1 (Architect) maps out every dependency and writes strict unit tests for each function * Agent #2 (Builder) writes the actual code to pass those tests * If a test fails, the Architect rejects the commit in ~350ms and forces the Builder to rewrite it * The loop doesn't break until the code passes every single check with zero errors left The result from a single run: 14 closed bugs, a fully rebuilt service layer, and zero hallucinations A dev agency would bill $6,000 and two weeks of revisions for this sprint. He got it in 9 minutes The trick is the closed loop: when one model writes and the other one questions every single line, the AI physically cannot invent fake methods or skip edge cases Same kind of adversarial agent setup that keeps showing up across the entire dev timeline right now The part nobody's posting is the exact prompt structure that holds this loop together. before someone packages it into an expensive paid tool Full breakdown and setup guide in the article below

Shoopy @0xShoopy

2 days ago

x.com/i/article/2113…

5 9 120 110K 276

Opus 5.5 got nerfed? TEST IT YOURSELF everyone's been posting about it for 4 days. almost nobody posts the receipt same prompt. 3 runs. one flag tells you who's guilty: → run 1: claude --bare -p "" --output-format json skips CLAUDE.md, hooks, skills, plugins, MCP and auto memory. that's the model alone → run 2: the same line without --bare that's the model plus everything your setup loads on every turn → run 3: /effort status, then /usage on Pro or Max it names what's eating your limit: subagents, skills, plugins, MCP servers compare total_cost_usd from run 1 and run 2. the gap isn't Opus. it's everything you loaded on top (--bare logs in with an API key, and both numbers are client-side estimates, not your bill) and it burns tokens faster? that gap is where they go. one Claude Code receipt every night. tonight's you run yourself.

Vinay @vinayp10

2 days ago

Claude, you're fired. I just moved my fiancée off Claude Code and onto @bot. We canceled her subscription. Since February, I've told everyone to run their second brain on Claude Code. This is the first time I've pointed someone at something else. x.com/i/article/2112…

32 15 602 177K 507

@noclipepe both true. the tools got free, the strategy part never will

THIS IS PURE F*CKING GOLD FOR GROK BOT TRADERS beyond the 1,894 public Grok Bot shares people pass around, there are 20 open-source trading repos with 622K+ GitHub stars you can point Grok Bot or Claude Code at. completely FREE. and most traders have never opened one full trading firms made of agents, bull and bear researchers who argue before every trade, risk teams that attack the decision, a foundation model trained on candlesticks, and engines that run the same strategy in backtest and live. all open source 20 GitHub projects worth opening first: 01 hypergrok-trading-desk → 7 role profiles and 17 skills for research, risk, execution and review on Hyperliquid. starts on public data with no wallet, runs in Grok Bot and Claude Code github.com/galleonlabs/hy… 02 awesome-grokbot → 1,894 live Grok Bot shares, every link status-checked and filterable. search it for a trading setup before writing your own instructions github.com/kydlikebtc/awe… 03 TradingAgents → a trading firm made of LLMs: 4 analysts, bull and bear researchers who debate, a trader, a risk team and a portfolio manager who approves or kills the call github.com/TauricResearch… 04 ai-hedge-fund → investor agents that run paper funds on real market days with fake money, plus backtests against a benchmark. works with Anthropic, xAI and more. never places a real trade github.com/virattt/ai-hed… 05 OpenBB → an open data platform for analysts, quants and AI agents. one place to pull market data before any agent forms an opinion github.com/OpenBB-finance… 06 alpaca-mcp-server → Alpaca's official MCP server. stocks, ETFs, crypto and options through plain language from Claude, Cursor or VS Code github.com/alpacahq/alpac… 07 Kronos → the first open-source foundation model for candlesticks, trained on data from 45+ global exchanges github.com/shiyu-coder/Kr… 08 qlib → Microsoft's AI-oriented quant platform: supervised learning, market dynamics modeling and RL, from first idea to production github.com/microsoft/qlib 09 RD-Agent → Microsoft's agent that automates the research loop. its quant version, R&D-Agent-Quant, was accepted to NeurIPS 2025 github.com/microsoft/RD-A… 10 FinGPT → open-source financial LLMs, including a forecaster and sentiment models released on Hugging Face github.com/AI4Finance-Fou… 11 freqtrade → a free crypto trading bot in Python with backtesting, plotting, money management and strategy optimization. control it from Telegram or a web UI github.com/freqtrade/freq… 12 hummingbot → automated strategies on centralized and decentralized exchanges. its Condor harness connects LLM decisions to deterministic execution, and paper mode needs no API keys github.com/hummingbot/hum… 13 ccxt → one trading API for 100+ crypto exchanges and prediction markets, in JavaScript, Python, C#, PHP, Go, Java and Rust github.com/ccxt/ccxt 14 nautilus_trader → a production-grade, Rust-native engine. research, deterministic simulation and live execution in one event-driven architecture github.com/nautechsystems… 15 Lean → QuantConnect's algorithmic trading engine. write the strategy in Python or C#, backtest it, then take the same code live github.com/QuantConnect/L… 16 jesse → research, backtest, optimize and live-trade your own crypto strategies. multiple timeframes and symbols without look-ahead bias, fully self-hosted github.com/jesse-ai/jesse 17 vectorbt → test thousands of trading ideas fast, then inspect trades, fees, drawdowns and portfolio performance in interactive charts github.com/polakowo/vecto… 18 backtesting.py → pip install, write one Strategy class, run one Backtest. the shortest path from an idea to a result on history github.com/kernc/backtest… 19 FinRL → the original reinforcement learning framework for trading. built for education, benchmarks and research prototypes github.com/AI4Finance-Fou… 20 machine-learning-for-trading → the code of the book, 3rd edition. 27 chapter guides from data sourcing to live execution github.com/stefan-jansen/… start with one repo you can actually read. give Grok Bot the link, have it read the README and run it on paper first read every agent's output and check one trade by hand before anything goes on a schedule then write the risk limits in yourself: position size, max loss per day, and a human yes before any real order 20 repos. 622K stars. $0 in software. none of them know your risk limits until you write them in a new stack like this every night the code is free. the tokens to run it through Grok Bot aren't. the article below shows 7 fixes that turn 7 usable results into 70 on the same budget ↓

Gresham @Greshamtrade

5 days ago

x.com/i/article/2112…

44 35 269 508K 881

@bonsaixbt start with one that has a paper mode. hummingbot runs paper without api keys

@SamgreyAi 5 kimi drafts for the price of one opus. iteration wins

@evgenia_vibes the remote finally lost its job to an agent

@slash1sol one layer per prompt is the whole trick. works for trading bots too

@0xlexlab 1,000 tries, one lucky curve. exactly

@0xKorens 100%. free code just means the tools are free. the edge is still on you

@0xVorzen paper first, limits always

@0xAIGOAT the smile asks for your inbox. the scope asks for your card

@VibeCoderOfek none, and that’s the point. the list is from the READMEs, not my PnL. nothing runs live without my limits

@Klonzu match rate over vibes is the whole point

@adib_builds boring and alive beats clever and liquidated

@0xVeyyro now point it at trading

@_Aquarelle bookmarks over likes usually means “i’ll try this sunday.” curious how many actually do

@0ai_zoyberg @sakatayasha @kepochnik curious which one holds up on a trading bot. pretty games are easy to judge, a backtest that quietly cheats isn’t

@quietoraclex checking /usage rn

@Richix01 and every fairy tale needs a villain. in backtests it's the 41 tickers nobody kept

GETTY IMAGES FILED FOR BANKRUPTCY YESTERDAY. THE PLAN CANCELS EVERY PUBLIC SHARE. YOUR F*CKING BACKTEST WILL NEVER KNOW IT EXISTED backtest audit, layer 2. this is how survivorship bias gets made: > chapter 11 in delaware, prepackaged plan > if it's confirmed, every class A share is cancelled. holders get nothing > part of the story: AI images eating stock photography a year from now GETY won't be on any ticker list. a backtest built on today's list skips it, and every one like it not rare. one study ran every US stock since 1926 (Bessembinder, 2018): > most lost to 1-month T-bills over their lifetime > about 4% of companies made all the net gains the quant course everyone's quoting this week ↓ says it straight: the dataset only holds companies still alive the split: → pull the universe as it was on each date, not as it is today → keep delisted tickers with their final price, even when it's zero → add delisting returns, not just the last close → rerun and compare against the "alive only" version you only step in three times: → when the dead list looks too short → when the edge shrinks after adding them. that's the real number → before anything goes near real money your backtest only met the winners. the market didn't paste this into Claude Code ↓ "Check my backtest in backtest/ for survivorship bias: 1. Show where the ticker universe comes from and when it was built. 2. Rebuild it point-in-time: for every rebalance date, only the stocks listed on that date. 3. Add delisted tickers with their final prices and delisting returns. Flag any I can't get data for. 4. Rerun both versions and show returns and sharpe side by side. 5. Paper only. Log every run to runs.csv. Show me the report first. Don't change any code until I say go." 7 layers. this was 2. tomorrow: the 0.2% fee that turns a $1,200 year into $600

Hanako @hanakoxbt

5 days ago

x.com/i/article/2111…

15 20 188 340K 540