🎁 $1,000 IN PLATFORM CREDITS🎁 One winner takes it. 📋 THE CHALLENGE Pick your weakest entry from any past comp. Read the feedback file we sent you. Rebuild it. Then prove the rebuild is better. Every entrant since the Cartographer has a file that says what worked, two fixes, and the one thing that would have moved them up. Most of those files have been read once. This week you act on one. This week's deliverable is the original pinned, the feedback file, the rebuild, and a control showing old against new. 🎯 PICK YOUR ENTRY Any comp you entered. Cartographer, Auditor, Translator, Control, or anything earlier if you still have the repo. Pick the one where the feedback file stung the most. That is the one with the most room. If you have never entered, you can still play. Pick any public repo from a past comp whose author has posted that it is open for rebuilds, and treat the public results post as your feedback file. Credit them in the README. A few sparks: - Your Cartographer map, rebuilt so the territory ships pinned in the repo - Your Auditor, rebuilt so every finding resolves to a provision a mod can open - Your Translator, rebuilt so the shape holds on the input that broke it - Your Control, rebuilt around the question the folder lost on - Any entry whose README a stranger could not walk The pattern is the same every time. The feedback named a gap. You close it and show the closure. 🔥 THE ANGLE THIS WEEK Last comp was The Control. You ran the same questions with and without a folder and committed both. The Rebuild runs the same test on two versions of the same folder. Original on one arm. Rebuild on the other. Same questions, same order. The questions come from the feedback file, so the test is aimed at the exact thing you were told to fix. In the Cartographer round, Adam put his README in front of an outsider, watched her fail for twenty minutes, cut it from 904 words to 471, and kept the old version in the repo so you could see the change. That is a rebuild with the receipt attached. This week everyone ships one.

After dealing with some rather restrictive NDA terms during a potential business collaboration, I started wondering: Why should protecting your ideas, projects and intellectual property be so complicated? Especially now that we're collaborating with AI agents, sharing Git repositories, developing software together and potentially turning projects into businesses. So I decided to put together something practical. NDA-UNIVERSAL-001 · Version 2.2 An English-language mutual NDA and non-use agreement, designed to protect both parties fairly. It covers: - Confidentiality and protection of existing intellectual property. - Independent development without unnecessary non-compete restrictions. - AI agents, external models and data processing. - GDPR and relevant US privacy considerations. - Git repositories, shared code and collaborative development. - Clear separation between access, collaboration and commercial exploitation. - Ownership of contributions and separately agreed commercial rights. I've also included a plain-English explanation for people who aren't lawyers, plus an ICM-compatible Markdown structure for AI-assisted workflows. Here's what I'd like to do next. Rather than keeping this as my own document, I'd love to explore turning it into a community-maintained standard. A version-controlled template where people can suggest improvements, identify unfair clauses and contribute practical experience. And ideally, have qualified legal professionals review it. Important: This is currently a draft based on German law, not a legally certified or universally enforceable agreement. The international aspects especially need professional review. I'm not looking for blind approval. I'm looking for constructive criticism. If something is missing, potentially unfair or legally questionable, let's identify it and improve it. And if anyone here is a lawyer, works with contracts or knows someone who might be interested in reviewing this, I'd genuinely appreciate the input.

Your AI says it’s done. The code looks good. The explanation sounds convincing. The agent reports success. But here’s the question nobody asks often enough: Did it actually finish the job, or did it just produce a convincing story about finishing the job? I’ve been thinking about this as a fundamental problem in AI system architecture. We spend enormous amounts of time improving prompts, models, memory and automation. But how much time do we spend verifying whether the result actually matches the original request? Here’s a simple rule I believe every AI agent should follow: No evidence = Not verified. Before an agent can declare a task complete, it should answer five questions: 1. WHAT exactly was accomplished? 2. WHERE is the actual result? 3. HOW was it tested? 4. WHAT failed or remains unverified? 5. CAN another agent independently reproduce or validate the result? If it cannot answer those questions, “completed” is a claim, not a verified status. And here’s the interesting part: This isn’t just about preventing hallucinations. It’s about building AI systems that remain reliable even when models change, agents are replaced, or projects are resumed months later. The model is temporary. The evidence should survive. Here’s a prompt you can try with your own agents: “Before declaring this task complete, separate verified results from assumptions. Provide the exact evidence, identify what has not been tested, and explain how an independent agent could verify your work. Never present an unverified claim as a completed task.” Now I’m curious: What’s the most convincing “TASK COMPLETED” message you’ve received from an AI that turned out to be completely wrong? I suspect we’ve all got a story.

Hi Everyone! I am sure this is a fraction of the capability some of you are all delivering, but wanted to share a win of mine, being a newbie and not fully advanced in the world of ICM. First Project of mine is - A Bookkeeping agent that collects Bank Statement CSVs via its Intray, reviews and processes these transactions with guidance and mapping, and delivering a CSV output to enter into my accounting software. This is for clients who don't use accounting systems such as Xero and Quickbooks for monetary reasons, or smaller businesses who want to save on their costs. I produced this project in around an hour and am so chuffed with results, it will save me hours when it comes to smaller bookkeeping and accounts preparation work. @Jake Van Clief thank you for introducing me to the world of ICM, I can't wait to see where this knowledge takes me, and I hope in the future I can build solutions for other businesses and perhaps profit from some!

I think we’re asking the wrong question about the future of work. Everyone keeps asking: “Will AI replace us?” But what if AI doesn’t simply replace human workers? What if it fundamentally changes who gives instructions and who follows them? Here’s my theory. Some people will learn to work with AI. Others will make AI work for them. There is a fundamental difference between learning how to use a tool and intuitively understanding how to design an entire system. You can teach someone prompting, workflows, automation and technical procedures. But following a procedure isn’t the same as understanding the underlying principles, recognizing patterns, connecting seemingly unrelated ideas and creating something entirely new. Imagine a future where an AI system coordinates operations, distributes tasks, monitors performance and tells human workers what needs to happen next. For some people, AI becomes their operational supervisor. For others, AI becomes their workforce. They define the vision, design the architecture, question the results and develop solutions that didn’t exist before. Some follow instructions. Others create the systems that generate those instructions. That doesn’t make one person more valuable than another. And these aren’t fixed categories. People can develop, learn and change roles. But it raises an uncomfortable question: Are we preparing people to operate within AI systems, or to think independently enough to build and direct them? And there’s an even bigger question. If AI eventually performs most economically productive work, who controls the resulting wealth? Who benefits from automation? Because the future of work won’t just be about intelligence or technical skills. It will also be about autonomy, ownership and the distribution of economic value. My prediction: The biggest advantage won’t belong to those who know the most prompts. It will belong to those who can recognize patterns, think in systems and turn original ideas into working solutions.