Daans 💎⚡ retweeted
THE WINNER OF ANTHROPIC'S OFFICIAL HACKATHON HAS PUBLISHED HIS CLAUDE CODE SETUP AND IT'S PURE GOLD
68 subagents, 286 skills, 94 commands, MIT license.
ECC turns Claude Code into a full engineering team.
The plan comes before the code.
The failing test comes before the fix.
Every change is reviewed by a context that hasn't seen how it was written.
→ Planning: you give it a sentence and approve the plan before a single line exists
→ Review: a different reviewer per language that reads your diff from scratch
→ Build fixes: a fixer per toolchain, including PyTorch and CUDA
→ Security: full OWASP sweep and injection scanning in the agent config
→ Architecture: detects design errors when they're still cheap, before migration
→ Domain: databases, ML pipelines, E2E tests, and documentation
The security part is what almost nobody sets up.
An external OWASP audit costs four figures and a week. This one is ready before you eat.
The skills cover testing, Python, Go, Rust, Django, Next.js, Docker, and CI/CD.
That said, don't activate all 286 at once or you'll end up with a config worse than the one you had.
This is how it integrates well:
→ Make a fork and read the README before copying anything
→ Start only with the planning agent and the rules pack for your main language
→ Copy to your .claude folder only the agents and skills you're going to use this week
→ Add the code reviewer and the security one afterward, as they yield the most
→ Test it first on a small repo, not in production
→ If something's extra, cut it. If you miss something, add just one more piece
The rule is simple: if you don't know what an agent is for, you don't need it yet.
Whoever integrates it well this weekend will spend the next quarter reviewing work instead of writing it.
Save this post, you'll thank me 🔖
HOLY F*CK... MIT students made $26 MILLION in ONE live round at a Citadel & Jane Street contest... from a single Python notebook that's on GitHub for free
3rd place, one notebook, the model Goldman Sachs built in 1990: Black-Litterman. no deep learning, no GPUs, just MATH
i ran their exact rule on 9 real stocks since October & let Claude Fable 5.2 make every call: the paper book is up 2.1% while SPY dropped 7.5%
first understand how this works:
> every team got the same stocks, live prices & 3 analysts' price predictions, then the contest scored the portfolios live
> Black-Litterman starts from what the market already holds & tilts toward the analysts, only as far as the math trusts them
> a Sharpe gate keeps a new portfolio only when it beats the current one, & the bot finished the live round $26 MILLION up
how to run it with Claude Code:
1. tell Claude Code to clone github.com/aardisaputra/uchi… & open case3.ipynb
2. have it run the notebook on their case3data first (it was written for Colab, so Claude swaps the Google Drive paths for local ones)
3. ask Claude to swap in real stocks & real analyst price targets as the views, Yahoo has both
4. keep their Sharpe gate, so a new target only goes in when it beats the current book
5. run it once a day after the close & log every target to a csv
6. let Claude decide each rebalance, execute or hold, with a reason (that's what Fable 5.2 does in the video)
7. paper trade the book & compare it with SPY every week
8. cap any single short, the raw rule sometimes asks for 89% net short
9. move to a live account only after a few months of paper results you like
i won't say anything more:
Daans 💎⚡ retweeted
A TRADER ASKED ME: "HOW DO I KNOW A TREND IS STILL ALIVE?"
I drew him one staircase
Higher high → BOS
Higher low → Hold
Higher high → BOS
Higher low → Hold
"As long as the lows hold and the highs keep breaking, it's alive."
He deleted half his indicators
Daans 💎⚡ retweeted
THE DAY OPEN IS WHERE THE TRAP IS SET
Open → The reference
Accumulation → Build the range
Manipulation → Break above the open
MSS → Break below the range
Entry → FVG, OB, or breaker
Close → Far from the open
Price goes above the open first
Then it goes where it was always going
More in this article:
Daans 💎⚡ retweeted
Hedge funds won't like this 33-page PDF.
Quants spend years stacking indicators into complex trend models.
This paper tested them against each other on 30+ years of data.
The winner: one simple EMA.
Adding more signals didn't beat it. It just made it easier to fool yourself with a backtest.
Drop the PDF into Claude, have it strip your strategy down to one EMA, then plug it into your Grok trading bot.
Here's the whole paper in 24 seconds:
Daans 💎⚡ retweeted
今天 GitHub 上刚冒出来一个猛货:OpenComputer(github.com/diggerhq/opencomp…)。
一句话介绍:它是天价 AI 程序员 Devin 的 100% 免费开源平替。
它直接给你的大模型配了一台完整的虚拟电脑沙盒,不用担心被AI 删库,直接把任务丢给它,它自己在后台虚拟电脑里写代码、跑终端、修报错。
大厂想卖几百美金一个月的全自动 AI 员工,开源社区三天就给你做成免费玩具。
F*CK YES... MIT students found the CHEAT CODE to Jane Street's trading challenge: an $8,000,000 P&L in one hour from 51 lines of Python
the cheat code: never guess direction - quote $1 under & $1 over the last trade & take a cut of EVERY trade, about 5,000 a round
it's called market making, the same game Jane Street plays for real
first understand how this works:
> Jane Street gave every team a fake exchange: 7 symbols, rival bots & thousands of trades in every 5-minute round
> the bot takes the last traded price as fair value & quotes $1 under & $1 over it, so its orders fill first
> it went from $300 a round with arbitrage to over $10,000 a round with this, & the final hour closed at $8,000,000
how to run it with Claude Code:
1. tell Claude Code to clone github.com/EnriqueKhai/JaneS… & hummingbot/hummingbot
2. have it read market_maker.py: the whole winning strategy is 51 lines
3. ask Claude to port it into a Hummingbot script: fair price = the last trade, one buy just under it, one sell just over it, cancel & replace on every new trade
4. start it on binance_paper_trade with BTC-USDT: live Binance data, no API keys
5. let it run a full session & save every fill to a csv, the one thing the winners said they wish they'd done
6. ask Claude to add an inventory cap, so the bot stops buying when it holds too much
7. have it widen the spread when BTC gets volatile: the challenge's bots were calm & real markets aren't
8. compare its P&L per hour with Hummingbot's own simple_pmm script
9. move to a live account only when a week of paper fills holds up
i won't say anything more:
Daans 💎⚡ retweeted
fu*cking every Grok Bot alternative
hft.md/grok-free
$0. No GPU. unlimited use
Daans 💎⚡ retweeted
Some guy made a python library of self-improving algorithmic trading AI agents
Then gave it away for free:
Daans 💎⚡ retweeted
4 TIMEFRAMES. 4 QUESTIONS. 1 TRADE.
4H → Which way?
1H → Where’s the liquidity?
15m → Did structure shift?
5m → Now place the trade
Skip one question
Lose the trade...
Daans 💎⚡ retweeted
🚨BREAKING: New Python Library for Algorithmic Trading with AI
It's called Ziplime (*not* Zipline).
This is what you need to know:
Daans 💎⚡ retweeted
🚨 مايكروسوفت تطلق إطار مفتوح المصدر لتشغيل نماذج بمئة مليار معامل على المعالج العادي! 💻🧠
الشركة أطلقت bitnet.cpp، إطار استنتاج لنماذج اللغة ذات البت الواحد (1-bit).
النتيجة المعلنة:
تشغيل نماذج بمئة مليار معامل محلياً على المعالج، من دون بطاقة رسوميات
استنتاج أسرع بـ6.17 مرة
استهلاك طاقة أقل بنسبة 82.2% على المعالجات
والكود مفتوح المصدر 100% على GitHub.
نماذج ضخمة كانت تحتاج خوادم وبطاقات غالية، صارت تعمل على حاسوب عادي. احفظ الأداة الان. 🧵👇
Daans 💎⚡ retweeted
لمن يرسم المخططات الهندسية ويبحث عن بديل مفتوح المصدر لـ AutoCAD
CADCraft برنامج رسم هندسي بلغة Rust يعمل بأسلوب AutoCAD: سطر الأوامر، ونقاط الالتقاط، والطبقات، والأبعاد، والتهشير، والبلوكات.
حسب المشروع، يفتح ملفات DXF وDWG ويحفظها، ويصدّر PDF وSVG وPNG. ويعمل على ماك وويندوز ولينكس وفي المتصفح، دون حساب أو اشتراك.
Daans 💎⚡ retweeted
L’IA est en train de révolutionner le design d’intérieur.
Nano Banana 2 peut désormais transformer un plan dessiné à la main en un rendu 3D en 4K avec des dimensions précises, générer des photos de chaque pièce et changer les meubles en un seul clic.
Ce qui coûtait autrefois 100 000 $ et prenait des mois ne coûte maintenant que quelques centimes et prend quelques minutes.
Voici un tutoriel étape par étape sur OpenArt :
Le genre de thread que tu vas regretter de ne pas avoir mis en signet. 🔖
🤝 Paid partnership · 🤖 Made with AI
Daans 💎⚡ retweeted
AI can generate a residential building worth hundreds of thousands of dollars in seconds.
This is 78 weeks of construction compressed into just 78 seconds.
A 10-story residential building. 80 apartments, 3,111 m³ of concrete, and 95.5 tons of steel reinforcement, all mapped into one construction sequence.
And this is where reality gets expensive.
You can watch every floor rise, every major component fall into place, and an empty plot become an entire residential complex.
Meanwhile, the internet is obsessed with AI-generated buildings that look incredible but explain absolutely nothing about how they're supposed to be built.
Turns out, generating the dream was the easy part.
Daans 💎⚡ retweeted
🚨 مطوّر بنى أداة واحدة، وصعد بها إلى المركز الأول على GitHub في يوم واحد!
اسمها REA: Reverse Engineer Anything.
أداة مفتوحة المصدر، برخصة MIT، تربط وكيل الذكاء الاصطناعي بمجموعة أدوات الهندسة العكسية. ترى ميزة تعجبك في أي تطبيق، حتى من دون كود المصدر، فتطلب من الوكيل أن يفهم كيف تعمل حتى مستوى الملف الثنائي، ويعرض الدليل، ثم يبني نسخة مكيّفة لمشروعك.
تحلّل البرامج الأصلية، وتطبيقات JavaScript وElectron، وتجميعات .NET، والمواقع. والتحليل يجري محلياً على جهازك.
رابط الأداة 🧵👇
Daans 💎⚡ retweeted
GPT 6 Astra + GitHub is getting really dangerous…
7 open source repos are enough to build a serious coding system around it
01 OpenHands
▸ github.com/OpenHands/OpenHan…
→ gives Astra an environment for editing code, running commands and working through real engineering tasks
02 Serena
▸ github.com/oraios/serena
→ understands symbols, references and repo structure before the model starts changing files
03 Context7
▸ github.com/upstash/context7
→ pulls current docs and examples when the task depends on external libraries
04 GitHub MCP Server
▸ github.com/github/github-mcp…
→ gives the agent direct access to issues, repos and GitHub actions
05 E2B
▸ github.com/e2b-dev/E2B
→ runs generated code and commands inside an isolated environment
06 pytest
▸ github.com/pytest-dev/pytest
→ turns the existing test suite into a hard gate before completion
07 Langfuse
▸ github.com/langfuse/langfuse
→ shows exactly what the model saw, called and broke during the run
what I’d actually give each layer
repo understanding → Serena
current docs → Context7
engineering workspace → OpenHands
GitHub access → GitHub MCP
runtime → E2B
verification → pytest
debugging the agent itself → Langfuse
GPT 6 Astra sits above all of it and handles the decisions that actually need reasoning
That is a much better use of a frontier model than making it waste context rediscovering files, reading stale docs and guessing why its own code failed
Daans 💎⚡ retweeted
Todo el mundo está haciendo Ingeniería Inversa de programas y videojuegos con Inteligencia Artificial.
Esta herramienta para tus agentes te ayuda a hacerlo.
Se llama REA. Un MCP open source para Claude Code, Codex y Cursor: github.com/morluto/rea
AI agents are getting f...cking dangerous with real data
9 GitHub repos for building a GPT 6 Astra system that can explore databases, write SQL, run analysis, and turn raw numbers into decisions
01 Vanna
▸ github.com/vanna-ai/vanna
→ natural language questions into SQL queries
GIVE AI ACCESS TO DATA
02 DB-GPT
▸ github.com/eosphoros-ai/DB-G…
→ build AI applications around private databases
03 MindsDB
▸ github.com/mindsdb/mindsdb
→ connect AI workflows to enterprise data
MAKE THE DATA USABLE
04 DuckDB
▸ github.com/duckdb/duckdb
→ run analytical SQL directly on local files
05 Polars
▸ github.com/pola-rs/polars
→ process large datasets with fast DataFrame operations
06 DataFusion
▸ github.com/apache/datafusion
→ build custom SQL query engines
TURN ANALYSIS INTO ACTION
07 PandasAI
▸ github.com/sinaptik-ai/panda…
→ ask questions about datasets in natural language
08 Evidence
▸ github.com/evidence-dev/evid…
→ turn SQL queries into interactive reports
09 Metabase
▸ github.com/metabase/metabase
→ explore metrics and build dashboards
the architecture:
question → schema → SQL → query → validate → analyze → visualize → decision
3 builds worth stealing:
finance: Vanna → DuckDB → Evidence
business: MindsDB → Polars → Metabase
research: DB-GPT → DataFusion → PandasAI
here's what makes this interesting
most AI agents can explain your business
very few can actually inspect the numbers behind it
GPT 6 Astra + the right GitHub stack changes what you can build
the next useful AI assistant won't just answer questions
it'll know where to look for the answer ⭣
Daans 💎⚡ retweeted
🚨 ACABAN DE FILTRAR UN TRADUCTOR CON IA QUE LE GANA A TODOS Y ES GRATIS
Lo hizo Tencent y funciona sin internet: pesa 440 MB y cabe en tu celular.
Es open source, cualquiera puede descargarlo, usarlo e integrarlo en sus apps.
Se llama Hy-MT2.
Lo que trae:
→ 33 idiomas, incluido el español
→ Traduce 100% en tu dispositivo, incluso en modo avión
→ Tus textos no salen de tu teléfono
→ Supera a la API de traducción de Microsoft en las pruebas de Tencent
Sin suscripción. Sin mandar tus datos a ningún servidor.
Te dejo el repo abajo 👇































