David A.

Claude's Stop hook can make it retry a task when a check fails, but the hook must exit with code 2. The difference matters: - exit 2 sends Claude back to the task and passes the hook's stderr output so it can fix the error. - exit 1 doesn't block completion, so Claude finishes even if the check failed. Exit 1 can also indicate a hook error, such as a missing dependency or a bug in the script. To trigger a retry, use exit 2 and send error details to stderr: [your command] >&2 || { echo failed; exit 2; }

OpenAI has released the source code for Apps SDK UI, making it easier to build ChatGPT plugins without creating an interface from scratch. The React and Tailwind library includes ready-made components: - Buttons and chat elements - Tooltips - Color themes - Responsive layouts It's designed specifically for building interactive interfaces directly inside ChatGPT. github.com/openai/apps-sdk-u…

Cloudflare has open-sourced Kitesurf, a browser built specifically for AI agents. Written in Rust and WebAssembly, it runs without a graphical interface, doesn't rely on Chromium, and can operate inside Cloudflare Workers. Kitesurf uses up to 7x less CPU and RAM than Chrome, but runs roughly 2x slower and cannot bypass CAPTCHAs. It's well suited for simpler agent tasks, such as taking screenshots, basic web scraping, and internet searches, and can work alongside Chrome.

Vercel has released UI SDK, allowing AI agents to interact directly with any interface built with it. For example, changing a project name, team members, and notification settings takes an agent just 16 seconds instead of 41 seconds with a traditional UI. According to Vercel’s official benchmark, execution cost is also more than 2x lower. Install the official skill to get started: npx skills add vercel-labs/ui --skill ui-sdk

All DeepSeek app features are now available through the API. 🔎 Search, file and link handling 🔒 Security features 🎙️ Voice capabilities Enable built-in features with just a few parameters, or use dedicated APIs for OCR, ASR, and TTS. You can also build more advanced chat interfaces or recreate the full DeepSeek experience in your own app. DeepSeek has also released an open-source chat demo supporting conversation history imports: cdn.deepseek.com/chat-demo

Claude Code’s desktop app now lets you create custom themes using mods! 🎧 Customize colors, add animated backgrounds, and tailor the interface to your taste. Just describe your desired look to Claude or use the theme-examples plugin. Update to the latest app version to try it out. Themes currently work only in local sessions. Cloud session support is coming later.

Thinking Machines has released the weights for Inkling 2 Small, a 594B-parameter MoE model with 17B active parameters. Pretraining took just 68 hours using 12T tokens. RL required another 9 days on a 2,048-GPU cluster. The model is competitive with Chinese models from 2–3 months ago. Running it requires: - 2× B300 or 4× H200 GPUs. Now waiting for a larger version.

Matt Pocock shared his Home Repo pattern for working with AI agents across multiple repositories. The idea is simple: create one central repository containing all skills, agent instruction files, wikis, and tools. Run agents directly from this repo and use `AGENTS.md` to point them to other projects: For skills, clone mattpocock/skills to ~/repos/skills. For Sandcastle, clone mattpocock/sandcastle to ~/repos/sandcastle This makes it possible to: - Keep all instructions and tools in one place. - Launch every AI agent from a single repository. - Make large-scale changes across multiple projects. A simple but highly effective approach. Matt has also published a detailed guide: aihero.dev/the-home-repo-pat… It includes a large collection of prompts and a `grill me` command that lets the agent ask clarifying questions, helping you adapt and improve the pattern for your own needs.

Claude Meetings brings Claude into team calls as an active participant, creating documents, interactive whiteboards, and other materials alongside the team in real time. 🎧 Team members can interact with Claude by voice or chat, request documents, and edit its work as it's being created. Claude also takes notes. After the meeting, its Claude page includes: - Discussion recordings - Action items - Full transcript - All generated materials Google Meet is supported at launch, with other video platforms coming soon. The beta is gradually rolling out to Team and Enterprise users.

Jensen Huang is stepping down as CEO of Nvidia. The company’s CFO will serve as interim CEO while the board of directors searches for a permanent successor.

Google is back in the game. The team has just released Gemini 4 Hydrogen, a new model with promising results. - Built on Carbon and optimized for high speed. - Outperforms Sonnet and Terra on benchmarks. - Scores 57% on Terminal-Bench 4.1. - Pricing: $1.50 per 1M input tokens and $7.50 per 1M output tokens. Currently available only to Pro and Ultra subscribers.

Just found a useful resource for vibe designers: DesignBookmark. It features 2,200+ design tools, websites, and resources across 59 categories. A new Motion section has just been added, featuring 76 animated components you can preview before adding to your project. Each component comes with a ready-made prompt. Click Claude or Codex on the card, copy the prompt, and let AI recreate the component in your project. designbookmark.com/motion

Chrome has updated its Modern Web Guidance skills to version 1.0. They give AI up-to-date knowledge of modern Web APIs, web development best practices, browser compatibility, and more—helping it avoid outdated approaches and use modern web platform features. The update also adds legacy code analysis. The project was developed in collaboration with the Firefox and Edge teams to ensure cross-browser compatibility. developer.chrome.com/docs/mo… github.com/GoogleChrome/mode…

Google now has its own Jev: Gemini Decisions. It's multimodal, accepting video, audio, images, and text as input. Pricing starts at $0.10 per million input tokens, and you can try it for free in Google AI Studio.

MiniMax has launched H3-Image, an image generator with native 4K output (4096×4096) in just 6 steps. It was built from MiniMax's video model H3 and supports prompts up to 32,000 characters and up to 16 reference images. Generation speed for 4K text-to-image: - Highspeed: ~10.3 seconds - Standard: ~16 seconds - Nano Banana 2.1 (Thinking Medium): 50 seconds in the same test Standard API generation costs ~$0.07, with a 50% launch discount. Until November 23, fal also offers 3–5 free generations. Under the hood, H3-Image uses a 33.1B diffusion transformer. The full model, including encoder and VAE, requires ~135 GB of downloads. SGLang supports offloading layers to RAM for 24 GB GPUs, and ComfyUI templates are already available. A two-step LoRA has also been developed, though its quality still needs testing.

OpenAI has launched a $200 retro console for vibe coders: the ModRetro Chromatic Codex Edition. It plays Game Boy and Game Boy Color cartridges and comes with a rewritable cartridge for your own games. With the ModRetro plugin, you can describe a game in Codex and have it build levels and animate sprites. In a demo, Codex created a snowboarding level that ran on the actual console. You can test games in the built-in emulator, then upload the finished ROM to a cartridge. This requires a firmware update and developer mode. OpenAI also released a browser-playable game gallery featuring titles like FlapGPT, with ROM downloads available.

OpenAI is turning ChatGPT into a new customer acquisition channel for businesses. Ads inside ChatGPT can now lead users directly to a company's AI agent instead of its website. Businesses will be able to: - Verify their business profiles and keep product information up to date. - Create custom AI agents to answer customer questions (for eligible US businesses). - Promote those agents through ChatGPT ads. Users click an ad and start chatting with the company's AI agent directly inside ChatGPT. OpenAI is gradually turning ChatGPT from a search tool into a platform where businesses can reach potential customers directly.

Gemini 4 Carbon is a beast. Using the new /swarm feature in Google Antigravity, a team of Gemini 4 Carbon agents turned historical photos and engineering blueprints into an interactive Three.js 3D exhibition — all from a single prompt. The project recreates the construction of the Eiffel Tower between 1887 and 1889. 😳

Alibaba has opened access to Qwen4-Max-Preview, a preview of its new Qwen 4 model. You can try it now through: - Qwen Chat - Agent Mode - Alibaba Token Plan The model targets complex coding tasks, office work, interface creation, and professional software development. No benchmark results are available yet. 👀 chat.qwen.ai/