标签：#X

Welcome back to TechCrunch Mobility — your central hub for news and insights on the future of transportation.

Part 1: Self-hosting on Jetson Orin Nano 👽 Jetson Orin Nano Web Server Follow-up...

Having spent over 25 years in software development and managing countless Linux environments, I've accumulated a vast collection of custom bash scripts, containers, and CLI tools. Remembering their exact paths and managing them efficiently directly from the terminal is a common challenge. To solve this, I built mytuis . mytuis is a small, attractive terminal UI for managing a personal catalogue of applications. It is built with gum and plain bash, with persistent storage in a human-readable YAML file. GITHUB REPO : https://github.com/horaciod/mytuis Why mytuis? I wanted a tool that didn't require heavy dependencies or a complex setup, but still looked great and provided a smooth user experience. Here is what mytuis brings to the terminal: CRUD operations: You can create, read, update, and delete application entries from a single menu. Quick launch: Pick an app from the filterable list and it is launched immediately. It replaces the manager process via exec, meaning no extra shell window is left behind. Smart path handling: It accepts absolute paths (like /usr/bin/firefox), relative paths (./scripts/myscript.sh), tilde paths (~/bin/foo), or plain command names looked up in your $PATH (firefox). Persistent metadata: Every entry stores its name, description, absolute path, creation date, and last-used date. Friendly TUI: You get clear menus, color-coded messages, and clean borders, all powered by gum. Under the Hood: Plain Text and Standard Utils Simplicity and standard compliance were key goals. mytuis requires bash ≥ 4 and standard Unix utilities like awk, sed, grep, date, and tput. Your catalogue is stored at ~/.mytuis.yaml. Because it is a standard YAML file, it can be inspected, edited, or backed up with any text editor. It is also completely safe to sync with a dotfiles repository or version-control. To ensure data integrity, all file operations are performed atomically by rewriting the YAML file from scratch on every change, so there is no risk of leaving the fi

A scan node sits at the leaf of the tree and pulls rows from a single table. A join node sits in the middle and brings together the rows that its two children send up. It takes one row from users , one row from orders , checks whether they belong to the same user, and if they match, emits the combined row. PostgreSQL has three nodes for this one job: NestLoop, HashJoin, and MergeJoin. The reason a single task splits into three nodes is much like the reason scans did. There is more than one way to find matching pairs from two inputs, and which way is cheapest depends on the size of the inputs and the shape of the join condition. Deciding which way is cheapest, by costing the alternatives, was the planner's job in an earlier chapter. This section looks at what those three nodes actually do when they execute. Given the same two tables, the three find matches in completely different ways, and that difference in approach is exactly what tells them apart. How the three nodes route requests All three join nodes are internal nodes with two children. One child is called the outer, the other the inner. All three run on the Volcano model's pull framework: when the parent asks for the next row, the join node takes rows from its two children, builds one matched row, and sends it up. The only difference is the order and manner in which it routes pull requests to its two children. NestLoop pulls the inner from the start all over again for each outer row it receives. HashJoin slurps the inner in one pass to build an index in memory, then takes outer rows one at a time and probes that index. MergeJoin, on the assumption that both sides are sorted in the same order, advances both sides one step at a time in lockstep. NestLoop: rescan the inner for every outer row The simplest method is NestLoop. As the name says, the loops are nested. The outer loop takes one row from the outer; the inner loop scans the inner from beginning to end, looking for inner rows that match that outer row. Wh

By the time 1.4 ends, the planner has produced one PlannedStmt. Inside it is an execution tree built from Plan nodes, frozen into a form you can follow step by step, something like "go into the primary key index on users, fetch the one matching row, then output that whole row." But that is still only a blueprint. Reading actual pages off disk, picking out the rows that match the condition, handing results back to the caller: none of that has happened yet. The stage that takes that blueprint and produces actual rows is the executor. The difference between the planner and the executor is the difference between deciding and doing. The planner was the stage that weighed "which index, in what order, with what join method" by cost and chose . The executor takes the chosen approach and carries it out as is . There is nothing left to choose. It just runs the nodes baked into the plan tree and pulls rows out of them. To run it, the executor takes the Plan tree it received and turns it into a PlanState tree. The Plan tree is the static blueprint the planner made, and it does not change during execution. But to actually run, each node needs state that changes as execution proceeds: which row it is reading now, whether the hash table is fully built, what tuple it has buffered from a child. So when execution begins, a PlanState tree with the exact same shape as the Plan tree is created. The blueprint Plan tree is left untouched, and the running state lives in that PlanState tree instead. How the executor produces result rows is the heart of the stage. The executor does not build the entire result set at once and stack it up. Instead, it asks the topmost node of the tree for "the next row," and that request travels down the tree to the leaves. When a leaf scan node reads one row from a page and passes it up to its parent, that row climbs up one level at a time through joins and filters until it reaches the top. The top sends that single row to the caller (the client, or the targe

A few weeks ago, I started exploring the Afriex MCP server. The setup was surprisingly straightforward. Connect your MCP client. Configure your API key. Verify the connection. Done. But then I ran into a different problem. Not a technical problem. A prompt problem. The Blank Prompt Problem Once everything was connected, I found myself staring at an empty prompt box. What should I ask? Sure, I could retrieve balances. I could create customers. I could generate virtual accounts. But what were the most useful workflows? What were the prompts that would actually help developers build real products? This isn't a problem unique to Afriex. It's becoming a common challenge across the entire MCP ecosystem. The infrastructure exists. The tools work. But many developers don't know where to start. MCP Changes How We Build Traditionally, integrating a payment API looked something like this: Read documentation Find the endpoint Write HTTP requests Parse responses Build business logic With MCP, the workflow looks very different. You can simply tell your AI assistant what you want to build. For example: Create a customer onboarding flow that: - Collects customer details - Generates a virtual account - Displays payment instructions Build it using Next.js and TypeScript. Instead of manually stitching everything together, the AI can interact with infrastructure through the MCP server. That's incredibly powerful. But only if you know what to ask. The Idea That's what led me to build the: Afriex MCP Prompt Cookbook A collection of practical, production-oriented prompts designed specifically for developers building with Afriex MCP. The goal is simple: Copy. Paste. Build. Instead of starting from scratch every time. The cookbook is open source and available on GitHub: https://github.com/SonOfUri/afriex-mcp-cookbook Feel free to explore the prompts, use them in your own projects, and contribute new recipes. What's Inside The cookbook is organized around real-world use cases. Not API endpoi

Originally published on MRTD.NET — fast, sourced news on crypto security, cyber & SEO. The uncomfortable first lesson You built a clean site, submitted a sitemap, maybe pinged IndexNow — and Google still shows nothing. Here's the part most guides skip: getting indexed by Google and getting indexed by everything else are two different problems , and conflating them wastes weeks. We separate what actually moves Google in 2026 from the folklore that just feels productive. Bing, Yandex and ChatGPT are the easy half If you've set up IndexNow , you've largely solved discovery for Bing, Yandex, Naver, Seznam and Yep — you POST your new/changed URLs to one endpoint and they get notified instantly. And because ChatGPT Search retrieves from Bing's index , confirmed Bing indexing effectively gates your visibility in ChatGPT's web results. That's a big chunk of the modern search surface handled with one integration. The catch: Google does not use IndexNow. It has said so repeatedly. So every "instant indexing" claim that leans on IndexNow is talking about Bing's world, not Google's. For Google, you need different levers. What actually gets you into Google There are really only two fast paths, plus one slow one. 1. Google Search Console — the only direct lever. Verify your domain (a private DNS TXT record; it does not trigger penalties or "re-evaluation," a common fear), submit your sitemap.xml , then use URL Inspection → Request Indexing on your key pages. There's a soft daily cap (~10–12 URLs), so spread a new site's pages over a few days. GSC is also the only place you can see whether a domain carries an inherited problem — essential if you bought an aged or expired domain. 2. Links on pages Google already re-crawls hourly. Googlebot's crawl budget for a brand-new, zero-authority domain is tiny. The fastest way to get a new URL discovered is a link to it from a page Google visits constantly — Reddit, Hacker News, Medium, established communities. These links are usually nofoll

Doing the work and being able to explain the work are two different skills. I've had the first one for 8 years. I'm building the second one now. I'm a Cloud Platform Engineer. AWS, Kubernetes, Terraform, Linux. Regulated environments, healthcare, production systems. Real experience. Almost zero public documentation of it. That's the gap I'm closing, starting from Day 1. The platform is KodeKloud. Each session gives you tasks across multiple tools. I'm posting the Linux and AWS tasks here. Here's what I built and what actually matters about each one. Task 1 (Linux): Create a User with a Non-Interactive Shell The task was to create a system user that can own processes but cannot log in interactively. This is what you do for service accounts. bash ssh user@hostname sudo su - useradd username -s /sbin/nologin cat /etc/passwd | grep username The /sbin/nologin shell is the important part. The user exists in the system, can own files and run processes, but cannot open a shell session. You verify by checking /etc/passwd because the last field in each line is the assigned shell. I've been doing this in production environments for years. I still verify every time. Not because I'm unsure. Because in a regulated environment, you don't assume, you confirm. Task 2 (AWS): Create an EC2 Key Pair via CLI aws ec2 create-key-pair \ --key-name my-key-pair \ --key-type rsa \ --key-format pem \ --query "KeyMaterial" \ --output text > my-key-pair.pem aws ec2 describe-key-pairs --key-names my-key-pair The private key is returned exactly once at creation. AWS does not store it. If you lose it, you generate a new one and replace it everywhere it was used. I've seen this cause real problems in production environments where the key wasn't backed up properly. Always run chmod 400 my-key-pair.pem after saving it. SSH will refuse to use a key file with open permissions. It won't tell you that's the reason straight away. What Day 1 Taught Me That 8 Years Didn't Nothing here was technically new to

SpaceX's AI1 satellite spans 70 meters tip-to-tip — wider than a Boeing 747 — and it exists entirely to run AI inference in low Earth orbit. Elon Musk posted the reveal video to X on June 9, 2026, ahead of SpaceX's IPO, with a three-word summary: "much simpler than Starlink." Each satellite produces 150 kW of peak AI compute and 120 kW sustained. SpaceX's roadmap calls for 1 GW of orbital AI compute capacity by late 2027, which at 150 kW per satellite means manufacturing roughly 6,700 AI1 units per year. To hit that number, they are building an 11-million-square-foot facility in Bastrop, Texas called Gigasat — nearly twice the floor area of Tesla's Gigafactory Nevada, dedicated to satellite production. The question is not whether the engineering works. SpaceX has launched more than 7,000 Starlink satellites. The question is whether orbital AI compute makes economic sense at scale, and that question nobody has answered publicly yet. The Reveal Wasn't Accidental SpaceX filed for its IPO at approximately $75 billion valuation in early June 2026. Musk's June 9 reveal of AI1 arrived within days of that filing. Orbital AI compute is the narrative SpaceX needs to justify a valuation that goes beyond launching satellites for other people. Every terrestrial cloud provider — AWS, Google Cloud, Azure — is competing for land, power, and cooling capacity to support the next generation of frontier AI. Musk's pitch is that those three constraints don't exist in space. The physics backs him up. The economics remain unproven. Why Space Has Structural Advantages for AI Compute The AI1 satellite's design exploits two physical realities that are impossible to replicate on Earth. Power is essentially free. In a sun-synchronous LEO orbit, a satellite receives near-constant solar illumination. SpaceX's solar arrays achieve 250 W/m² power density without atmospheric attenuation. The marginal cost of electricity after the capital investment in the array is close to zero — no grid contracts,

Hello Dev Community! 👋 It is officially Day 50 — a massive half-century milestone on my daily, unbroken streak toward mastering full-stack MERN engineering! Reaching Day 50 feels absolutely incredible. Yesterday, I mapped out dynamic path parameters. Today, I wired the input engine by building a complete asset workflow: Capturing Host "Add New Product" data payloads and committing them to local file storage pipelines! Following Prashant Sir's backend sequence , today was all about bridging the gap between host client forms and backend architecture using the Model-View-Controller framework. 🧠 Key Learnings From Day 50 (Product Ingestion & Storage) Processing data mutations sent from input forms requires tight coordination between parsing middlewares and file serialization engines. Here is how I structured the logic today: 1. Intercepting Form Submissions ( POST /host/add-product ) Set up a clean route mapping inside hostRouter.js to process dynamic data blocks sent by the host. The endpoint parses input parameters securely via backend streams. 2. Utilizing Class Instances for Storage Instead of directly pushing raw unstructured dictionaries into file records, I initialized a new object instance using my Day 48 structural class framework ( new houseList(...) ). This forces incoming data attributes—like name, price, location, and images—to match my exact system layout blueprint. 3. Asynchronous File Serialization Invoked the instance method .save() , which runs a non-blocking background task: it reads the active database layout array inside homesdata.json , appends the newly formulated object safely, and flushes the stringified update back onto the hard drive array using Node's fs operations. javascript // A conceptual look at how my controller hands data over to the model layer today const Product = require("../model/home"); exports.postAddProduct = (req, res) => { const { title, price, location, rating, imageUrl } = req.body; // Instantiating the core class data mold

Hello Dev Community! 👋 It is officially Day 48 of my unbroken full-stack engineering journey! Yesterday, I refactored my modular core patterns into MVC architecture. Today, I linked up a major functional extension inside the /model layer by introducing JavaScript Classes (OOP) to coordinate my local file operations and storage data patterns! Instead of writing loose object definitions, I stepped up my enterprise game by structuring a reusable class footprint that encapsulates data parameters and handles non-blocking file-system persistence asynchronously. 🧠 Key Learnings From Day 48 (OOP Modeling & File Systems) As clearly shown in my development workspace layout within "Screenshot (116).png" , modeling data with dedicated classes shifts your core structural logic from simple scripts into highly scalable engines: 1. The Model Data Blueprinter ( constructor ) I used the standard ES6 class framework inside home.js to structure an explicit data mold ( houseList ) with attributes tracking: houseName , price , location , rating , and photoUrl . This ensures every entry traveling through our server follows an identical structure. 2. Streamlining Async Persistence ( save() ) Rather than relying on globally declared floating arrays, my .save() blueprint method triggers an internal lookup to read existing data stacks asynchronously before safely using fs.writeFile() to serialize and flash mutated JSON rows into a local data asset ( homesdata.json ). 3. Static Decoupled Fetchers ( static fetchAll() ) I mastered using static methods. Since reading a data grid requires pulling records without creating an instance of a single house first, making fetchAll(callback) static allows our controllers to tap the hard disk records straight from the class reference layout: javascript // A conceptual look at my file-reading design today static fetchAll(callback) { const filePath = path.join(rootDir, 'data', 'homesdata.json'); fs.readFile(filePath, (err, data) => { if (err) { callback(JSON.

Đối với một developer, trackpad không chỉ là thiết bị điều hướng mà còn là công cụ tối ưu hóa workflow. Khi làm việc với các IDE nặng như VS Code hay IntelliJ, khả năng phản hồi của trackpad quyết định tốc độ xử lý tác vụ. Trong bài so sánh giữa ROG Zephyrus G14 GA403 hay MSI Creator 16 AI? Đâu là lựa chọn cho sáng tạo chuyên nghiệp? , trải nghiệm trackpad là một điểm nhấn quan trọng. Trải nghiệm cử chỉ và độ chính xác trong lập trình Khi làm việc với code, các cử chỉ như chuyển đổi desktop ảo (Virtual Desktops) là cực kỳ quan trọng để tách biệt môi trường chạy Docker, trình duyệt và editor. Vuốt 3-4 ngón: Cả hai dòng máy đều hỗ trợ tốt, nhưng trên MSI Creator 16 với diện tích lớn hơn, việc nhận diện cử chỉ vuốt ngang giữa các workspace mượt mà hơn đáng kể. Độ chính xác chọn văn bản: Với một developer, việc bôi đen một đoạn code dài hoặc chọn chính xác một ký tự nhỏ là yếu tố sống còn. Trackpad trên G14 có độ nhạy cao nhờ kích thước gọn nhẹ, trong khi Creator 16 cho cảm giác vững chãi, ít bị trượt hơn khi thao tác nhanh. Độ trễ (Latency): Cả hai đều đạt chuẩn cao, tuy nhiên trên Windows, trải nghiệm đôi khi không mượt bằng macOS. Để khắc phục, việc sử dụng driver tùy chỉnh là cần thiết. So sánh hệ điều hành và mẹo cấu hình cho Developer Trải nghiệm trackpad thay đổi rõ rệt giữa Windows và Linux : Windows: Hỗ trợ tốt Precision Drivers. Bạn nên vào Settings > Bluetooth & devices > Touchpad để tinh chỉnh độ nhạy.\n- Linux: Nếu bạn dùng Ubuntu hay Fedora, hãy cài đặt libinput . Để tối ưu hóa cho workflow code, bạn có thể cấu hình file .wslconfig nếu chạy môi trường Windows Subsystem for Linux nhằm đảm bảo tài nguyên không bị nghẽn khi thao tác giao diện.\n Thông số kỹ thuật tóm tắt: ROG Zephyrus G14 GA403: Ryzen 9 8945HS, RTX 4070, 32GB LPDDR5X, OLED 14" 120Hz, nặng 1,5 kg. MSI Creator 16 AI Studio: Core Ultra 9 185H, RTX 4080/4090, lên đến 64GB DDR5, Mini LED 16" 120Hz, nặng 2,1-2,5 kg. Bài viết này là bản tóm tắt kỹ thuật. Xem chi tiết tại bài gốc.

Most agents drag their entire past into every turn. A better default: keep a thin index of what was said hot, and fetch only the few turns you actually need — intact, on demand. Code: github.com/NirajPandey05/jit_context There is a quiet assumption baked into how most agents handle memory: that more context is safer than less. If the model might need something, put it in the window. The conversation grows, every prior turn rides along on every new request, and we trust the model to find the part that matters. That assumption breaks twice. It breaks on cost , because an agent loop re-sends its whole window on every step — a hundred stale turns aren't paid for once, they're paid for on turn 101, 102, and every step after. And it breaks on quality , because models don't read a long window evenly. Relevant facts buried in the middle get underweighted; irrelevant bulk competes for attention with the thing that actually answers the question. Past a point, a bigger context produces a worse answer, not just a costlier one. So the interesting question isn't "how do we fit more in?" It's "how do we keep the window small and dense without losing the one old turn that matters?" This post is the design we built around that question — for the specific case of long conversation history — plus the benchmark we used to keep ourselves honest. 01 · The mechanism: a hot index over a cold store The design borrows directly from how computers have always managed memory that doesn't fit: a small fast tier that's always present, a large slow tier that holds the bulk, and a rule for moving things between them. Virtual memory pages between RAM and disk. We page between the context window and an external store — for attention instead of address space. Concretely, there are two tiers. The cold store holds every turn at full fidelity, keyed by id — nothing is thrown away. The hot index holds one compact entry per turn: a short summary, a little metadata (entities, whether the turn recorded a dec

Call it a startup with a sole founder and a very large seed round, but what's next is less clear.

As a solo developer I wanted a fast, mobile-friendly interactive map for Palworld that didn't bury me in ads. The result is Pindrop , and here are a few of the technical decisions behind it. Rendering 1000+ markers without jank The interactive map uses Leaflet with a custom marker-clustering layer. Markers are served as static JSON from the edge and hydrated client-side, so the first paint is server-rendered and the heavy marker work happens after. A breeding calculator as a pure function Palworld's breeding combos are deterministic, so the breeding calculator is just a lookup over a precomputed table rather than a backend call. That keeps it instant and fully cacheable. Stack Next.js (App Router) for SSR + static generation Leaflet for the map layer Supabase for the small amount of dynamic data Vercel for hosting and edge caching If you play Palworld, the guides section collects the breeding, location and boss notes I kept losing track of. Feedback from other devs welcome — especially on the clustering approach.

Hi everyone 👋 I'm a Full Stack Developer with experience in JavaScript, React.js, Next.js, Node.js, Express.js, MongoDB, and WordPress development. Over the last few years, I have worked on multiple projects ranging from company websites to full-stack web applications. In this article, I want to share my experience transitioning from the traditional MERN stack to Next.js and why it has become my preferred framework for modern web development. Why I Started with MERN Stack The MERN stack was my first choice because it allowed me to build complete applications using JavaScript. Technologies MongoDB Express.js React.js Node.js Benefits ✅ Single language across frontend and backend ✅ Huge ecosystem ✅ Fast development ✅ Easy API integration Challenges I Faced As projects became larger, I started facing issues like: SEO limitations Performance optimization Routing complexity Code organization Server-side rendering requirements This is where Next.js entered the picture. Why I Switched to Next.js Next.js provides several powerful features out of the box. Server-Side Rendering (SSR) Pages can be rendered on the server which improves SEO and initial page load performance. 2.** Static Site Generation (SSG)** Perfect for blogs, landing pages, and marketing websites. 3.** App Router** The new App Router makes routing cleaner and more scalable. Server Components Less JavaScript is sent to the browser, improving performance. Better Developer Experience Features like: File-based routing Built-in image optimization Middleware API routes make development faster and cleaner. My Current Tech Stack Frontend Next.js React.js TypeScript Tailwind CSS Backend Node.js Express.js MongoDB Tools Git & GitHub Postman Vercel Contentful CMS What I Learned The biggest lesson I learned is: Focus on solving real business problems instead of chasing every new technology. Frameworks will change, but understanding JavaScript fundamentals, APIs, databases, authentication, and system design will always be

Là một developer thường xuyên đối mặt với các lỗi memory leak trong hệ thống C++/Rust, việc chọn môi trường profiling là yếu tố sống còn. Nếu bạn đang cân nhắc giữa việc chạy Valgrind hay Heaptrack trên WSL2 so với Native Linux (hoặc máy trạm có RAM SO-DIMM nâng cấp được), đây là những trải nghiệm thực tế từ quá trình debugging. Hiệu năng Valgrind và Heaptrack: Sự khác biệt rõ rệt Khi sử dụng valgrind --tool=memcheck , tốc độ thực thi thường giảm xuống còn 10-50 lần so với bình thường. Trên Native Linux , việc quản lý bộ nhớ diễn ra trực tiếp trên kernel, giúp các công cụ này hoạt động ổn định nhất. Ngược lại, trên WSL2 , do lớp ảo hóa và cơ chế quản lý memory của Microsoft, bạn sẽ thấy overhead đáng kể hơn. Đặc biệt là khi tạo Heaptrack flame graph , việc phân tích bộ nhớ lớn có thể khiến WSL2 bị giới hạn bởi file .wslconfig nếu không cấu hình đủ RAM.\n Lệnh thực thi nhanh: # Chạy Valgrind trên hệ thống của bạn valgrind --leak-check = full --show-leak-kinds = all ./your_app # Sử dụng Heaptrack để lấy flame graph chi tiết hơn heaptrack ./your_app WSL2 Overhead và bài toán phần cứng (Onboard vs SO-DIMM) Một vấn đề thực tế là khi profiling các ứng dụng nặng, bộ nhớ hệ thống bị chiếm dụng cực nhanh. Nếu bạn đang dùng laptop với RAM onboard 16GB , việc chạy đồng thời Docker + IDE + Valgrind trên WSL2 dễ dàng dẫn đến tình trạng swap liên tục do giới hạn cứng của phần cứng. Từ kinh nghiệm thực tế, nếu công việc yêu cầu profiling chuyên sâu thường xuyên, một chiếc máy có RAM SO-DIMM cho phép nâng cấp lên 32GB hoặc 64GB sẽ là cứu cánh tuyệt vời. Bạn có thể tham khảo thêm về sự khác biệt giữa ReviewLaptop để hiểu rõ tại sao việc chọn đúng loại RAM lại quan trọng cho workflow của một developer. Bảng so sánh nhanh: | Đặc điểm | Native Linux | WSL2 (Ubuntu) | --- | --- | --- | | Speed Overhead | Thấp hơn (Direct Kernel) | Memory Management | Trực tiếp, ổn định | Có lớp ảo hóa, dễ bị giới hạn bởi .wslconfig | | Flame Graph Rendering | Mượt mà | Đôi khi chậm do I/O file qua hệ th

Telegram argues India should block specific content, not an entire platform used by millions.

DeductiveAI, a startup that uses AI to catch and resolve bugs in software, was founded just three years ago.

Your repo is already your agents' context, whether you designed it on purpose or not That sentence took me a while to understand. In this post I'll save you the trip. It was October 2025, working in Skyward's monorepo with AI agents every day. And every day the same routine: I'd tell the agent in the prompt "don't use this", "don't do it this way", "reuse the component that already exists". I wrote it down. I repeated it. The agent did exactly what I told it not to do. It wasn't that it didn't listen to me. It was that it read the code and saw something else there. The agent believes the code, not your prompt An agent follows the patterns it sees in the repo, not the ones you tell it in the prompt. And subagents are worse, because they start without the conversation's context. The whole fight you put up earlier in the chat, for them it never happened. So this is what kept happening. It created a new component even though one already existed that solved exactly that problem. It didn't respect the design rules or use the design tokens. It followed stale docs because they were still there, alive, with nothing flagging them as outdated. My first instinct was everyone's instinct, cram more context into the prompt. More rules, more "please don't do this", more examples pasted in by hand. It half worked, and for the next task you had to add it all again. Until the next subagent showed up and started from scratch. At some point, tired of repeating myself, I understood the obvious thing. The agent wasn't disobeying me. It was reading the repo and listening to what the repo said about itself. If the good component lives alongside three old versions, it has no way to know which one is the official one. If the docs say one thing and the code does another, it'll believe whichever is closest at hand. It's doing exactly what I asked. The repo itself is the context agents use. If it's badly structured, the answers won't be good. Period. No prompt fixes a repo that contradicts itsel