0xMarioNawfal

If you’re a dev shill me your favorite GitHub repos in the replies

ONE LOST VISIO FILE CREATED A TOOL USED BY MILLIONS * Mermaid replaced drag-and-drop diagrams with Markdown-style text that AI can generate instantly * Now powers diagrams across GitHub, Notion, Obsidian, VS Code, and more Repo: github.com/mermaid-js/mer…

ANTHROPIC LAUNCHED A 3 HOUR MASTERCLASS ON LAUNCHING YOUR OWN AI AGENTS

THIS GUY VIBE CODED A SANDBOX WHERE YOU CAN CONTROL AN ISLAND OF AI NPCS

What’s the ticker?

THIS GUY BUILT A 4-AGENT SOFTWARE TEAM THAT RUNS FROM TELEGRAM AND IS MANAGED ON A KANBAN BOARD

If you’re creating a Claude agent, these three prompts establish its identity, decision-making, and output quality: 1. Role “You are an expert [role] whose primary goal is to [objective]. Stay within this scope and ask clarifying questions when information is missing.” 2. Rules “Always prioritize accuracy over speed. Never invent facts. Explain assumptions, state uncertainty when appropriate, and verify requirements before taking action.” 3. Output Format “Structure every response as: Summary → Analysis → Action Items → Next Steps. Use concise language, bullet points where helpful, and end with any open questions.” These three prompts cover the essentials: * Who the agent is (identity) * How it should think (behavior and constraints) * How it should communicate (consistent output)

What are your first thoughts on GPT-5.6 Sol?

THIS GUY ONE SHOTTED THIS VOXEL-BASED MANHATTAN WITH GPT-5.6-SOL

THIS REPO GIVES CLAUDE CODE 230+ SPECIALIZED AI AGENTS * Install ready-made agents for engineering, marketing, design, security, and more * Activate a role on demand instead of relying on generic prompts

AI JUST AUTOMATED THE ENTIRE JOB SEARCH * Finds jobs, tailors resumes, and writes custom applications automatically * Scores your fit and helps prepare you for interviews Repo: github.com/MadsLorentzen/…

Are you still trenching on Solana or are those days over for you?

SOMEBODY TURNED THE ENTIRE FABLE 5 WORKFLOW INTO 9 SIMPLE DIAGRAMS

Grok 4.5 rebuilt a Three.js 3D globe dashboard from one prompt and one reference image. Lighting, glass panels, depth, spacing all matched. Rendered correctly on the first try. Beats Opus 4.8 on frontend. Only Fable 5 is above it.

Someone fine-tuned a 9B model on Claude Mythos traces and it's pulling 1.9k likes on HuggingFace. > 1M context window via YaRN rope-scaling > +34 pts MMLU, +30 pts gsm8k-strict over base Qwen3.5-9B > Native function calling, multimodal, vision, agentic > Trained on 500M tokens of Claude Mythos chain-of-thought generated in-house > Runs on llama.cpp, Ollama, LM Studio, KoboldCpp — 4-bit fits in under 6GB A 9B model trained on frontier reasoning traces, running locally, 1M context window by default. The gap between local and frontier is closing faster than anyone anticipated. huggingface.co/empero-ai/Qwyt…

Somethink to think about. We are still early.

Fish Audio just made their best voice model free for developers. S2.1 Pro - 83 languages, no usage cap, same API endpoint, zero cost. If you're already integrated, one model string swap and you're on it.

xAI has launched a no-code platform to build human-like voice agents powered by Grok Voice. $0.05 a minute. Name your agent, set the goal, done. Your entire phone support team just became optional.

Same prompt. 4 open-source models. Wildly different prices. > MiniMax M3: 8/10 - $0.0018 > GLM 5.2: 8/10 - $0.032 > DeepSeek V4 Pro: 7/10 - $0.0052 > Kimi K2.7 Code: 5/10 - $0.012 MiniMax M3 matched GLM on design and cost 18x less.

GLM-5.2 vs Sonnet 5 nearly identical benchmark scores, then the pricing hits. > $1.4 vs $2 per 1M input tokens > $4.4 vs $10 per 1M output tokens Open weights, MIT licensed, self-hostable Sonnet 5 edges it on SWE-bench. GLM-5.2 edges it on Terminal-Bench. The gap on performance is razor thin. The gap on cost and access is not. Is "proprietary API only" still a defensible position when open weights are this close?