AIDotDev Digest #001

Hey devs! Welcome to AIDotDev. Your weekly AI newsletter built for developers. We cut through the noise, bringing you the latest trends, tools, and projects to keep you ahead. No AI anxiety, just pure growth. Let’s build the future together!

Author

AIDotDev
March 16, 2025

Inside today’s AI.DEV newsletter

📰 AI DEV BITES

  • OpenAI Agent SDK Unleashed 

    OpenAI’s new Agent SDK drastically reduces AI agent development time from weeks to minutes with an easy-to-use Python toolkit. It automates task sequencing, tool integration, and coordination between specialized agents, eliminating the need for complex code. Built-in safeguards reduce errors like hallucinations, and tracing makes debugging faster.

    Why it matters: You can quickly prototype AI applications with minimal effort on a platform designed to streamline agent workflows.

  • Windsurf 4 Speeds Coding 

    Codeium’s Windsurf Wave 4 enhances your coding experience. Cascade now provides live previews of code changes, while auto-linting fixes errors as you type, reducing manual cleanup. Tab-to-Jump suggests imports instantly, and suggested actions anticipate your next move in the codebase. Powered by Claude 3.7, it’s faster and more intuitive.

    Why it matters: Developers can work more efficiently with fewer keystrokes, speeding up iterations on large projects.

  • Gemma 3 Powers Mobile AI 

    Google’s Gemma 3 brings powerful AI to phones and single GPUs, no server required. The 1B model (529MB) generates text at 2,585 tokens per second, ideal for offline mobile apps like note summarizers. The 27B version, with a 128K-token context, outperforms Llama-405B (Elo 1338) for code and document analysis on a single NVIDIA GPU. Fine-tune it on Hugging Face to debug code or support 140 languages. It can even generate image captions on-device.

    Why it matters: Build fast, local AI tools with PyTorch or Keras—your next app could launch sooner.

  • Figma x Bolts Boost Design 

    Bolt.new’s new collaboration with Figma, powered by a partnership with AnimaApp, allows users to transform Figma designs into pixel-perfect, production-ready full-stack applications with ease.

    Why it matters: This streamlines design-to-development handoffs, enabling designers to bring ideas to life without full developer support. Note: The feature is live, visit bolt.new for details, including a free trial (3 conversions per month) and token-based pricing for extra use.

📚 CURATED LEARNING RESOURCES

  • What is MCP? by Alex Xu 

    Explore Anthropic’s Model Context Protocol (MCP) via Alex Xu’s X thread. Learn how it connects AI models like Claude to tools and data, with diagrams for building MCP servers, perfect for devs integrating AI into workflows.

  • Using LLMs for Code by Simon Willison 

    Simon Willison’s blog post shows how to leverage LLMs (like Claude) for coding, offering prompt tips and workflows. Ideal for developers wanting to boost productivity with AI, including real-world examples and Python tools.

  • A Single-File MCP Server, In TypeScript 

    Learn to build a single-file MCP server in TypeScript with no build step, as outlined in Matt Pocock’s guide. Perfect for devs integrating Anthropic’s MCP into apps, streamline AI tool connections with clean code.

  • Perplexity’s MCP Server for Claude 

    Check Perplexity’s MCP server on GitHub for Claude, linking it to real-time web research via Sonar API. Great for devs building AI tools needing internet access. Enhance Claude’s context with dynamic data.

  • Lovable x Anthropic x Supabase Build Comp 

    Join Lovable, Anthropic, and Supabase’s build competition on Lu.ma, create AI projects (likely MCP-based). Ideal for devs wanting hands-on AI experience, networking, and prizes, boost your skills fast.

  • Dario Amodei on U.S. AI Leadership 

    This Council on Foreign Relations talk reveals the Anthropic CEO's predictions about AI development, directly impacting every developer's career prospects. Amodei states that AI will write 90% of code within 3-6 months and possibly "almost all code" within a year, shifting developers' roles from coding to high-level planning and oversight. As frontline developers, understanding this imminent industry transformation is crucial for adjusting your career plans.

🛠️ WEEKEND DEEP TOOL: LLMS.TXT STANDARD

In this era of rapid AI development, Large Language Models (LLMs) have become an essential part of our digital lives. As Andrej Karpathy mentioned, we have entered the age of LLMs, but there's an important issue to consider: existing website documentation is almost exclusively designed for humans, not optimized for LLMs.

This is the background behind the birth of the llms.txt standard.

What is llms.txt?

llms.txt is a newly proposed website standard aimed at providing structured website information for LLMs through a specifically formatted Markdown file in a website's root directory. This standard was proposed by Jeremy Howard, to address the key limitations LLMs face when processing website content.

Why is llms.txt needed?

LLMs currently face two major challenges:

  1. Context window limitations: Most website content is too large to fit completely into an LLM's context window

  2. HTML conversion difficulties: Converting complex HTML pages (with navigation, ads, and JavaScript) into LLM-friendly plain text is both difficult and imprecise

llms.txt provides an elegant solution, allowing websites to offer more concise, expert-level information centralized in an easily accessible location.

llms.txt Specification and Structure

The llms.txt file uses Markdown formatting but follows a specific structural specification. It should include the following sections in this fixed order:

  1. H1 Title: The project or website name (the only required section)

  2. Blockquote: A brief summary of the project, containing key information necessary for understanding the rest of the file

  3. Detailed information: More detailed information about the project and how to interpret the provided files (optional)

  4. File lists: Sections delimited by H2 headers, containing URLs where further detail is available

Each "file list" consists of a Markdown list, with items containing a required Markdown hyperlink [name](url), optionally followed by a : and notes about the file.

Additionally, it's recommended to provide clean Markdown versions of pages that might be useful for LLMs by appending .md to the original page URL.

Practical Use Cases

The applications of llms.txt are extremely wide-ranging:

  • Software documentation: Helping developers and LLMs find API documentation and programming guides more quickly

  • Corporate websites: Outlining company structures and policies

  • E-commerce: Explaining products and shopping policies

  • Educational institutions: Providing quick access to course information and resources

  • Personal websites: Helping answer questions about personal resumes

How to Implement llms.txt

As a developer, you can add llms.txt support to your project or website through these steps:

  1. Create a Markdown file that conforms to the specification

  2. Place it at the /llms.txt path in your website's root directory

  3. Provide clean Markdown versions of important pages (by adding .md to the original URL)

It's worth noting that all fast.ai and Answer.AI software projects using nbdev already generate .md versions of pages by default.

Coexistence with Existing Standards

llms.txt is designed to coexist with current web standards:

  • Unlike sitemap.xml, it provides a curated overview for LLMs rather than a list of all pages

  • It can complement robots.txt by providing context for allowed content

  • It can reference structured data markup used on the site

Conclusion

As AI and LLM technologies evolve, optimizing content for these intelligent systems will become increasingly important. The llms.txt standard provides a simple yet effective way for websites to offer more valuable information to LLMs, enhancing their ability to understand and process information.

As a developer, implementing llms.txt can not only improve the quality of LLM interactions with your content but may also become a standard practice for websites in the future. After all, in the LLM era, making your content friendly not only to humans but also to AI will be an important competitive advantage.

To learn more or see examples of llms.txt, you can visit llmstxt.org or check the provided directories such as llmstxt.site and directory.llmstxt.cloud. Also you can see how Anthropic write llms.txt Anthropic & Model Context Protocol

⚙️ NEW TOOLS

  • 🗄️ Perplexity’s MCP Server
    Real-time web data for Claude via GitHub’s MCP server. Fork it, tweak the Sonar API, and ship dynamic AI tools fast.

  • 🤖 Google Gemma 3
    Run AI locally on phones/GPUs. Perfect for offline apps, vision tools, and fine-tuning across 140 languages.

  • 🖍 Figma x Bolt
    Figma-to-code in clicks at Bolt.new. Automate React/Vue apps from designs—less grunt work, more building.

  • 🌐 Nano Browser
    Open-source Chrome extension on GitHub. Hook your LLM to automate web tasks locally, private, no cloud bloat.

  • 🔑 Blender MCP
    Anthropic's MCP for 3D modeling. Automate Blender tasks and analyze 3D data with AI assistance.

🧸 AI-Powered Building: How People Do It

Check out this impressive AI-powered Elden Ring recreation! A solo developer used Kamui AI for generating assets (backgrounds, characters), TripoAI for 3D modeling, Claude 3.7 for development assistance, and AI audio tools for sound, all integrated with Three.js.

The creator had never played Elden Ring and learned solely from watching livestreams. This showcases how AI tools are making complex game development accessible to individuals.

🔧 SMALL WEEKEND PROJECT : 3D MONET WITH BLENDER & MCP

Recreate Monet’s "Woman with a Parasol" in Blender with MCP
Love 3D modeling? This weekend, build a 3D version of Monet’s painting using Blender and Anthropic’s MCP. It’s perfect for devs exploring 3D or Three.js, export your model for web apps!

  • Setup (5 mins): Follow the Blender MCP setup guide to connect Claude AI. Use Claude to analyze your progress or suggest tweaks via screenshots.

  • Build (10 mins): Recreate Monet's "Woman with a Parasol" painting as a 3D scene: 

     
  1. Environment Setup:
     * Create a grass field plane using particle systems to generate tall grass and wildflowers
     * Set up a bright sky background with an environment texture to simulate Monet's blue sky with white clouds
     * Use volumetric lighting and scattering to mimic sunlight filtering through the air
  2. Character Modeling:
     * Model a female character wearing a loose white/light blue long dress
     * Add cloth physics to the dress to make it naturally flow in the "wind"
     * Create a small boy character and position him in the background
     * Design the pose of the female figure looking up at the sky with a sense of lightness
  3. Props Creation:
     * Model a green parasol with detailed textures and semi-transparent effects
     * Adjust the parasol angle to match the painting
  4. Materials and Textures:
     * Use procedural or hand-painted textures to simulate Impressionist brushstrokes
     * Add slight color oversaturation to capture the Impressionist color style
     * Consider using displacement maps to mimic oil painting texture
  5. Lighting Setup:
     * Set up a strong key light to simulate bright sunlight
     * Add blue ambient light to simulate sky reflections
     * You may need additional fill lights to soften shadows
  6. Render Settings:
     * Consider using NPR (Non-Photorealistic Rendering) techniques to simulate a painted effect
     * Add brushstroke effects in post-processing to enhance the Impressionist style
  You might also want to use volumetric scattering effects in Blender's EEVEE or Cycles renderer to capture the feeling of sunlight penetrating the air, which is an important element in Monet's work.

    Tip: If your render lacks color/texture, tell Claude, “This is my final render—why no color/texture?” (Send a screenshot for help).

Why it matters: This bridges 3D modeling, AI automation, and web dev export to Three.js for interactive web scenes. Share your 3D masterpiece on X with #AIDotDevProject!

Final :

Although there may still be some differences, this is just an example. By having more conversations with AI and providing better descriptions, it should be possible to achieve similar results—and even accomplish much more.

AI is evolving rapidly, and it's natural to feel overwhelmed. Every week, I'll share the key updates you need to know, helping you slow down in this fast-moving world and truly absorb knowledge you can apply. See you next week!

If you find this helpful, please share it! Your support is what keeps me motivated to write every week.

If you have any suggestions, feel free to email me: [email protected]