1. I use pen.dev (not the cursor extension, the desktop app), an amazing design tool that allows byok. With copilot and Claude I've had full access to the full suite of tools and multi-agent designing.

  2. Why Your AI Agent Can't Solve CTF Challenges: Pwnable Why Your AI Agent Can't Solve CTF Challenges: Pwnable An AI agent can find a vulnerability and write exploit code yet still fail to capture the flag. It might overwrite a function point…

  3. I used to keep a messy Notion page full of prompts that I’d copy-paste into new chats five times a day, one for editing, one for SEO outlines, one for rewriting stiff drafts. Finally organized them into proper custom instructions with clea…

  4. Subscribe to updates for Elevated errors on claude.ai, Claude Code, and platform.claude.com via email and/or text message. You'll receive email notifications when incidents are updated, and text message notifications whenever Claude create…

  5. Descobrir que o Opus 5.5 tem um cap de 330 carácteres e infere o resto com o contexto, sem indexação de código local isso faz aquelas perdas de contexto que te irritam tanto fazer mais sentido. Minha dúvida é, fable tem esse mesmo cap no a…

  6. TL;DR: Lookpack is a self-hosted training server + mcp that gives Claude desktop the data to become your running coach. An iOS app syncs your Apple health data (workouts, sleep, HRV, weight, nutrition) to a server you run yourself and a bu…

  7. I've been using Claude Code for small freelance jobs and wanted to share what keeps it from going wrong: Start every job with a read only tour. Ask it to explain the repo before it edits anything.

  8. I built a free, open-source skill for Claude Code (also works with Codex and Cursor) based on Karpathy's tip: ask LLMs to write "80% of the way to ASD-STE100", the controlled English used in aircraft maintenance manuals. **What it does** I…

  9. Hi HN, I'm Crokily. Pots is an AI coworker that lives in the chat your team already uses: Slack, Discord, Telegram, or any channel with a Chat SDK adapter.

  10. Live Threat Interception Sandbox Experience why probabilistic LLM guardrails fail and how deterministic capability boundaries block 100% of autonomous tool hijacking in sub-millisecond time. Attack Vector Generator Red Team AegisKernel Pro…

  11. Every new Claude Code session I re-explained the same things: why we picked this auth flow, which commands actually work, what I'm in the middle of. CLAUDE.md helped, but it turned into a 400-line dump that nobody, including Claude, read c…

  12. The scheduled prompt is the same every day: you’re in a coding contest, write one self-contained HTML file, no libraries, no network, you pick what to build, and it can’t repeat anything you’ve built before. No follow-ups, no fixes.

  13. As the title suggests, I am looking for suggestions as to how I can maximize or where to use my $200 monthly API credits for claude. Unfortunately it couldn't be used to generate images like codex API, and couldn't be used as credit to top…

  14. Has anyone else found Claude harder to work with lately, despite how capable it is? I use my own provider-agnostic orchestration layer, so I regularly switch between coding agents.

  15. Was using chatgpt for the last 4ish months, but with the super fast drain, and constant reset juggle, I have had enough, and wanted to try out Claude. I wanted it to make a 1v1 fighting game, with local and online pvp.

  16. Notes on model training Musings on long-context fine-tuning Notes on training a 27B vision-language model on long-context examples. On this page - Getting one example to train - Sharing the repeated work I recently fine-tuned Qwen3.6-27B,…

  17. I built Furnish for the Kane CLI Hackathon. It's an AI-powered home design studio where you can generate editable floor plans, furnish rooms, make changes using AI, and explore your design in 3D.

  18. A couple of days ago, a new update for Claude desktop completely disabled its ability to push to GitHub and, as I reckon, to any git remote. That is a hard rule at the network layer, so it's unavoidable.

  19. I run a set of Claude conversations in one project. One of them, The General, is my chief of staff: it merges, publishes and tests overnight.

  20. Here I have 3 Christmas trees. I prompted each agent " built me a Christmas tree widget in 2 minutes " just a quick fun experiment.

  21. First timer in making a game. I am using Claude Code and Opus 5.5.

  22. I'm calling it Agents Overwatch and it's at https://agentsoverwatch.com/ I'm probably not using nearly as many agents as you guys, but I have about 12-15 I use on a regularly basis. All of them are grouped by project.

  23. Hello, I mainly run an e-commerce business and had started developing minor apps for the business successfully in the beginning of last year with the help of AI and several services. Small apps, not too complicated.

  24. Hi everyone, I use Claude Code via the Claude Desktop app because it’s by far the simplest and cleanest workflow for me. However, there is one really frustrating UX limitation: whenever images are displayed in the conversation (screenshots…

  25. 200 items

    Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2 is a post-trained model with 27 billion parameters, offering five safetensor tiers including BF16 and FP8 variants; it shows improved performance in stopping sequences without sacrificing other benchmarks like GPQA (down from 13 to 3 truncations). GPT-6 Astra has been showcased for its capabilities in AGI tasks such as skin cancer diagnosis and robotic manipulation, with scores of 62.7% on ARC-AGI-3 Se…

  26. Loft Your files. Your AI.

  27. DeepEP-Ascend (中文介绍)DeepEP-Ascend 是面向华为昇腾 NPU 的高性能机器学习训练与推理通信库,提供 MoE dispatch/combine 的专家并行(EP)all-to-all 操作,支持 FP8 dispatch 和延迟 epilogue。此外还提供流水线并行(PP)、面向上下文并行和数据并行(CP/DP)的 Bucket 集合通信,以及 Engram 远端内存访问等通信原语(开发中)。其公开 buffer API 与 NVIDIA 版…

  28. I keep hearing that AI now writes most of the code at some companies. For people actually shipping Python with Claude Code, Cursor or Copilot: Roughly what share of your merged code starts as AI output?

  29. Add rate limiting Add rate limiting to the upload endpoint and prove it holds. - 1ClaudeDraftAdds limit middleware with tests.6:58 AM - 2ChatGPTAgreesRan the unit and integration tests.6:59 AM - 3GeminiObjectionA burst of requests still ge…