What if a skill could improve itself? I built one. Then let it loose on its own source code for 10 rounds. The result: 52 versions. 109 experiments. 100 adversarial eval prompts conquered. Zero discards. (I'm not making this up, check in comments) Meet the Autoresearch Loop skill. Based on Karpathy's autoresearch methodology, it systematically improves anything measurable: other skills, n8n workflows, system prompts, SOPs, business processes. The loop is simple: Define "better" → Lock everything else → Change one thing → Test → Keep or discard → Repeat The execution is not. Why is it so good? We ran it against itself across 10 rounds. Each round we wrote 10 new, harder eval prompts targeting gaps the skill didn't cover yet. Each round it started near 0% and climbed to 100%. By the end it handles things like: • Emergency production hotfixes mid-loop • Multi-stakeholder metric conflicts • LLM-as-judge scoring drift • Artifact forking when scope grows too broad • Campaign cost/ROI tracking across months • When to abandon a loop entirely vs retire an artifact 462 lines. Every single one battle-tested against a prompt specifically designed to break it. This is the skill that makes other skills better. The proof is in the methodology: it improved itself 52 times, until we ran out of ways to challenge it. Download it at give it a try with any of your current skills!

🚨 MIT just dropped a game-changer for anyone building with AI (Full document attached) They analyzed 831 ways companies are making AI safer and sorted them into 4 buckets: - Governance & oversight - Technical & security - Operations - Transparency & accountability Operations dominates with 36% of all safety measures. What's everyone actually doing: - Testing & audits (127 different approaches) - Setting clear data rules + live model monitoring - Publishing risk assessments that buyers actually trust What almost NO ONE is doing (<1% adoption): - Model alignment checks - Conflict of interest shields - Whistleblower systems - Energy impact tracking This is a MASSIVE opportunity gap 👀 Why this matters for our community: 1. Regulators are watching - They're 100% using this list to craft new rules 2. Investors are asking - Show them you have these controls = faster funding 3. Early movers win - Lock in these practices now before they become mandatory https://airisk.mit.edu/blog/mapping-ai-risk-mitigations

For the last few months I have been compiling every single bit of information I could find about the "New" SEO for LLM Search Engines like ChatGPT or Perplexity (also called GEO, LEO, LLMO, LLM SEO or AIO) to understand how they work, how to make the most out of them and how to position websites so that they can be found at all. It's being not an easy task as the available information is still very scarce, but I think I already have enough to put it all together in a practical guide. Traditional SEO is not dead (yet), but it will be by 2027 and ignoring this now, could cause that no-one finds your site via organic searches or LLM searches any more very soon. This means that your web traffic will uniquely come from people who is accessing your site directly, or through links. And the numbers don't lie: 📊 80% of consumers now rely on AI-generated content for 40% of their searches. 📉 25% reduction in organic web traffic due to zero-click searches. And it's going down quickly. ⚡ 1,200% growth in generative AI traffic in just 8 months. 🎯 90% of ChatGPT citations come from pages ranking 21+ (NOT your precious top 5 rankings!). Here's the brutal truth: While some are still optimizing for Google's Page 1, ChatGPT is mining pages 3, 5, and 10 for answers that millions trust daily. The shift is massive: → ChatGPT processes 1+ billion messages daily → Google's market share dropped below 90% for the first time since 2015 → Semrush predicts LLM traffic will COMPLETELY overtake traditional search by 2027 Traditional SEO is NOT prepared for this. The mechanics between generative AI and classic search are fundamentally different. But here's the opportunity: Early adopters are seeing 35% increases in featured snippets and 27% more brand citations by implementing LLM optimization strategies.

I'm trying to download a file from a website using an LLM + the following MCPs - puppeteer, filesystem, fetch, and memory: https://pastebin.com/bgysAe39 The file may be PDF, image, etc. Suggestions for a better way to approach / MCP for me to experiment with? Thanks in advance! Response from the LLM: The fetch tool retrieved content from the URL, but it indicates that the content type is image/tiff and cannot be simplified to markdown. It returned the raw content, which appears to be binary data (likely the image data of the TIFF file). This confirms that the URL obtained from the puppeteer_evaluate step is indeed a direct link to a file. However, since the file is a TIFF image and not a text-based format, the fetch tool returned its raw binary content, which is not directly readable as text. While fetch successfully retrieved the file content, I cannot interpret or save it as a human-readable file like a PDF with the current tools. The file system tools can save text content, but they are not designed to handle arbitrary binary data like this TIFF image. Therefore, even though I managed to get the direct file URL, I cannot complete the task of downloading and saving the file in a usable format with the current capabilities. To summarize: 1. I successfully used puppeteer to navigate to the page with the sales information. 2. I used puppeteer_evaluate to find a link and extract a potential file URL. 3. I used fetch to retrieve content from that URL, confirming it's a TIFF image file. 4. I cannot save this binary image data as a file using the current file system tools. I have reached a point where I can access the file content's raw data, but I lack the ability to process and save this specific file type (TIFF image).

@Didac Fernandez and all, what are your recommendations for the best MCP servers for TypingMind, and why? What types of tasks have you been using them for?