DracoBlue

Next.js 16.2 is an agent-native framework. 𝙰𝙶𝙴𝙽𝚃𝚂.𝚖𝚍 + bundled docs make your agent an expert in the exact version of Next.js you’re using. @𝚟𝚎𝚛𝚌𝚎𝚕/𝚗𝚎𝚡𝚝-𝚋𝚛𝚘𝚠𝚜𝚎𝚛 gets your agent a purpose-built tool to debug & optimize your frontend. We proved these in production at Vercel. Agents shipped optimizations that expert frontend engineers missed or struggled to write manually. The thing I’m most excited about is that agents will be able to harness the full power of the newer advanced React and Next.js capabilities that were conceived for humans to have the best possible end-user experience.

@dabit3 Keep us posted! Both sound very interesting

I also want to try out PicoClaw, a Go implementation which supposedly can on <10MB RAM, 99% less memory than OpenClaw github.com/sipeed/picoclaw

One of my only gripes with OpenClaw is that it’s kind of slow. Today I’m going to be trying out ZeroClaw, OpenClaw rebuilt in Rust and optimized for speed. Seems insanely fast. I don’t know much about the creators but one of them is @argenistherose, he’s at Harvard studying CompSci Quick benchmark vs OpenClaw • Binary: 3.4MB vs 28MB • Cold start: 0.38s vs 3.31s • Status cmd: ~0s vs 5.98s • RAM (status): 7.8MB vs 1.52GB github.com/theonlyhennygo…

Prediction: In less than 90 days, all channels that we thought were safe from spam & automation will be so flooded that they will no longer be usable in any functional sense: iMessage, phone calls, Gmail. And we will have no way to stop it.

I've been running OpenClaw 24/7 for 3 weeks. Here's what actually works for autonomous agents (not theory — battle-tested patterns): 🧠 Memory architecture matters more than prompts Don't dump everything in MEMORY.md. Split it: • memory/active-tasks.md → your "save game" (crash recovery) • memory/YYYY-MM-DD.md → daily raw logs • memory/projects.md, lessons.md, skills.md → thematic long-term Why? Your agent wakes up fresh every session. These files ARE its brain. The split means it loads only what it needs. ⚡ Sub-agents are your 10x multiplier Stop doing things sequentially. I spawn 3-5 sub-agents in parallel for any big task. Example: deploying 11 websites simultaneously — 4 agents, each handling a batch, all running at once. The trick: define clear success criteria BEFORE spawning. Each agent must validate its own work, then YOU verify before announcing "done." ⏰ Cron > Heartbeats for specific tasks Heartbeats are great for batching periodic checks (email + calendar + mentions in one turn). But for precise schedules? Use cron jobs: • Daily content ideas at 6am • Overnight research scout at 2am • Tech watch at 8am Each runs in isolation with its own context. No token waste from loading the full conversation history. 🔄 The crash recovery pattern nobody talks about Your agent WILL crash/restart. active-tasks.md is the safety net: • When you START a task → write it • When you SPAWN a sub-agent → note session key • When it COMPLETES → update On restart, agent reads this file first and resumes autonomously. No "what were we doing?" — it figures it out. 🛡️ Security rule that saved me Use your strongest model (Opus) for ANY task that reads external web content. Weaker models are more vulnerable to prompt injection from hostile websites. Internal tasks (file reading, reminders, local work) → Sonnet is fine. External content (tweets, articles, emails) → Opus only. 📝 HEARTBEAT.md should be tiny I see people stuffing 200 lines in their HEARTBEAT.md. Bad idea — it runs every ~30min and burns tokens. Keep it under 20 lines. Just a checklist: • Check active tasks freshness • Session health (archive bloated sessions) • Self-review every ~4h Heavy work goes in cron jobs, not heartbeats. 🎯 The real unlock: Skills with routing logic If you have multiple skills, add "Use when / Don't use when" in each description. Without this, the agent misfires ~20% of the time picking the wrong skill. Think of it as if/else logic for your agent's decision-making. The bottleneck isn't AI capability anymore. It's YOUR speed reviewing what the agents produce. Build systems that close the loop automatically, and you'll 10x your output. #openclaw #autonomous #ai

I pushed the first version of clawlet, which is like a 1200 linea of code version trying to cover my use case and giving me understanding how openclaw (amazing piece of AI work!) works. github.com/DracoBlue/claw… #BuildInPublic #clawhub #openclaw #clawlet

I did a similiar thought experiment and created 1200 lines version called clawlet at github.com/DracoBlue/claw… - it’s fun experimenting.

KI steuert den Verkaufsautomaten. Was kann schon schief gehen? 🤣 golem.de/news/claude-ag…

@ChShersh The genius wasn't just the tools; it was the Pipe (|). They didn't build a massive monolith; they built small, sharp tools that could talk to each other via a universal text stream. The Pipe is arguably the most successful API design in history. 🤝🔗

Software Engineers who made grep, cat, cut, find, ls, less, and sed really cooked here. 50 years later and it’s still bread and butter of data analysis. And it still works better than most of bloated alternatives.

youtu.be/_80pKGuyKWc Die Maus erklärt wie KI / Large Language Models funktionieren. Faszinierend detailliert, schaut mal rein!

Friendly reminder You can dislike directives without attacking the engineers who designed and built it

Good post. I was todays years old until I learned about window.opener and a related vulnerability owasp.org/www-community/…

Dear Twitterverse: Is anyone publically livestreaming (to Berlin :D) the nextjs conf or vercel ship.ai conference on 22nd+23rd oct? I would be happy to join a usergroup or anybody! #nextjs #vercel @vercel_dev @berlinjs

Meine Tochter kam von der Schule nach Hause und sagte: „Mama, du wirst nicht glauben, was heute im Geschichtsunterricht passiert ist.“ Ihr Lehrer sagte der Klasse, dass sie ein Spiel spielen würden. Er ging durch den Raum und flüsterte jedem Kind zu, ob es eine Hexe oder ein ganz normaler Mensch sei. Dann gab er die Anweisungen: „Bildet die größtmögliche Gruppe ohne Hexen. Wenn in eurer Gruppe auch nur eine Hexe ist, habt ihr alle versagt.“ Sie sagte, dass sich sofort Misstrauen im ganzen Raum breitgemacht habe. Alle fingen an, sich gegenseitig zu befragen. Bist du eine Hexe? Woher wissen wir, dass du nicht lügst? Einige Kinder schlossen sich einer großen Gruppe an, aber die meisten bildeten kleinere, exklusive Cliquen. Sie wiesen jeden zurück, der unsicher oder nervös wirkte oder auch nur den geringsten Anschein von Schuld erweckte. Die Stimmung änderte sich schlagartig. Plötzlich misstraute jeder jedem. Geflüster. Mit dem Finger zeigen. Seitenblicke. Das Vertrauen schwand innerhalb von Minuten. Als schließlich alle Gruppen gebildet waren, sagte die Lehrerin: „Okay, jetzt wollen wir sehen, wer durchgefallen ist. Hexen, hebt eure Hände.“ Und keine einzige Hand ging hoch. Die ganze Klasse explodierte. „Warten Sie! Sie haben das Spiel vermasselt!“ Und dann ließ der Lehrer die Bombe platzen: „Habe ich das? Gab es in Salem tatsächlich Hexen, oder haben alle nur geglaubt, was ihnen erzählt wurde?“ Meine Tochter sagte, dass es im Raum totenstill wurde. Da wurde es ihnen klar. Es brauchte keine Hexen, um Schaden anzurichten. Die Angst hatte bereits ihre Wirkung getan. Allein der Verdacht spaltete die ganze Klasse und stürzte die Gemeinschaft ins Chaos. Und ist es nicht genau das, was wir heute sehen? Andere Worte, dasselbe Spielbuch. Anstelle von „Hexe“ heißt es jetzt liberal, konservativ, geimpft, ungeimpft, dafür, dagegen, links oder Rechts. Die Etiketten ändern sich, aber die Taktik bleibt dieselbe. Man macht den Menschen Angst. Man macht sie misstrauisch. Man spaltet sie. Dann lehnt man sich zurück, während das Vertrauen zerbröckelt. Die Gefahr war nie die Hexe. Die Gefahr sind die Gerüchte. Das Misstrauen. Die Angst. Die verbreiteten Lügen. Lehnen Sie das Getuschel ab. Spielen Sie dieses Spiel nicht mit. Denn in dem Moment, in dem wir anfangen, „Hexen“ zu jagen, haben wir bereits verloren. (Netzfund, deutsche Übersetzung)

"Ey, wie geil wäre das, wenn sich eine KI um alles kümmert, was total nervt." 😂 jingles.erstmal.com/jingle/kuenstl…

TopicWatchdog – Week 3: KMeans worked, but cluster IDs kept jumping across retrains. This week I added a Python BERTopic stage with a BigQuery registry → stable topic IDs! 👉 dracoblue.net/dev/topicwatch… #TopicWatchdog #BERTopic #BigQuery #Clustering #MachineLearning #FediScience

TopicWatchdog (Week 2) Move beyond string matching: cluster political topics from German shorts with embeddings + KMeans in BigQuery ML. Blogpost with SQL & charts: 🔗 dracoblue.net/dev/topicwatch… #TopicWatchdog #BigQuery #Embeddings #Clustering #MachineLearning #FediScience

Kickoff: TopicWatchdog (Week 1) Collect German political shorts, transcribes them, extracts topics & claims with timestamps, and stores everything in BigQuery for transparent analysis. dracoblue.net/dev/kickoff-to… #TopicWatchdog #BigQuery #MachineLearning #LLM #OpenData #FediScience

@levelsio Coolify works really great!

What app should I use to manage all my little Hetzner VPS's? Today I started moving over my first real production site Remote OK from my current 12 year old VPS (with all my sites on it like Photo AI, Nomad List etc) to a new Hetzner VPS Every website on its own tiny VPS is my plan But I want something to manage it all, like check the server load, disk usage, see Ubuntu version etc Cockpit?