H M Riyadh Ul Islam

Vedran B.

https://npgsqlrest.github.io/ • 19K followers

I barely survived this massive straw-man of an article. It started really promising: --- It’s very trendy to hate on Domain Driven Design (DDD). So trendy in fact, that disdain for DDD has united disjoint sets of overly opinionated developers. "Cracked" engineers don’t like it because all that squishy stuff about language and behavior adds noise to their data structures. "Systems Unicorns" hate it because all those a-words (abstractions) hurt data-alignment. The "Vibe Coders" never got it. Who needs that boomer talk when code is disposable? Claude will just write the whole thing over anyways. The zeitgeist has spoken. DDD is boring. It’s old, it’s slow. It has ED. --- Then it goes on to explain that DDD fights complexity because it does better code-file organization in a project by giving them context. Well, I hate DDD because it is a practice of implementation of (physical) models in database clients, instead of, well, you know, the database (where it actually belongs). They promote and enforce the concept of persistence ignorance, because a database is persistence and persistence is, you know, just a device, and we don't want to be bothered with those details when thinking big. So instead, we will implement our model in, well, memory (not a device at all). Because SQL bad, and OOP and Functional good. Can someone inform them that modern databases are already persistence ignorant (actual devices are abstracted), and all you are promoting is, well, the database ignorance? In any case, when we implement our physical model in a database client like memory (not a device at all) then we also have to have another physical model in the database, because, well, persistence, and then we have two models, instead of just one... so much simplicity, but anyway. Yes, I am one of those DDD haters. https://lnkd.in/dd9qmvHw

I barely survived this massive straw-man of an article. It started really promising: --- It’s very trendy to hate on Domain Driven Design (DDD). So trendy in fact, that disdain for DDD has united disjoint sets of overly opinionated developers. "Cracked" engineers don’t like it because all that squishy stuff about language and behavior adds noise to their data structures. "Systems Unicorns" hate it because all those a-words (abstractions) hurt data-alignment. The "Vibe Coders" never got it. Who needs that boomer talk when code is disposable? Claude will just write the whole thing over anyways. The zeitgeist has spoken. DDD is boring. It’s old, it’s slow. It has ED. --- Then it goes on to explain that DDD fights complexity because it does better code-file organization in a project by giving them context. Well, I hate DDD because it is a practice of implementation of (physical) models in database clients, instead of, well, you know, the database (where it actually belongs). They promote and enforce the concept of persistence ignorance, because a database is persistence and persistence is, you know, just a device, and we don't want to be bothered with those details when thinking big. So instead, we will implement our model in, well, memory (not a device at all). Because SQL bad, and OOP and Functional good. Can someone inform them that modern databases are already persistence ignorant (actual devices are abstracted), and all you are promoting is, well, the database ignorance? In any case, when we implement our physical model in a database client like memory (not a device at all) then we also have to have another physical model in the database, because, well, persistence, and then we have two models, instead of just one... so much simplicity, but anyway. Yes, I am one of those DDD haters. https://lnkd.in/dd9qmvHw

Charles Cozad

Pallium • 2K followers

Game changer: Claude Code paired with Playwright CLI Playwright allows computers to work with the visual hierarchy of web pages, interact with page elements and take screenshots. Other ways to interact with web pages from Claude include the Claude Chrome extension and the Playwright MCP server. Both options are heavy on token use because they force the LLM to process images and a large MCP surface area. With the Playwright CLI you get token efficient browser processing. By default it's headless but you can also have it show what is seen in a window. Some ways Claude + Playwright CLI helped me recently. Testing Some AI generated code was completely broken on some API calls. The problem was the data shape was wrong so the tests "passing" meant nothing. Adding Playwright in the loop provided Claude a way to test the code in the same way a user would use the app. Iterative Design I pointed Claude at the current state and the reference for a "good" design I wanted to move towards. I interacted in real time regarding how the page was looking and directed test cases to further check behavior. I am currently exploring other use cases and I'm excited for a functional level up in my workflow. Are you doing visual work with AI? What tools are you using?

Eric Frazier

OpenSource Brokers • 390 followers

https://lnkd.in/dDC5UXtt I made a wrong turn with my first MCP. The bottom line? In spite of the many articles and tools I have found to estimate tokens, to compress, like using XML instead of JSON and other tricks. MCP is not for moving data. It says pretty clearly, it is intended for: MCP provides a standardized way for applications to: - Share contextual information with language models - Expose tools and capabilities to AI systems - Build composable integrations and workflows On that note, I think Claude is really crappy for suggesting using Pagination when it hits that 25K MCP message size limit, because that just leads down a very stupid and very slow in-efficient path. Like 7 API calls to get 255 rows of csv!! And it takes like a minute or more to fetch what would be milliseconds with a DB query. It looks like my own example with Claude Code and this 25K token limit per message isn't always the case, but increasing that limit isn't cheap. You can increase the limit a little bit though in your .claude/settings.json $ cat settings.json { "env": { "MAX_MCP_OUTPUT_TOKENS": "50000" }, "hooks": { "SessionStart": [ { But one thing I clued in on, is that Claude hooks themselves don't pass large amounts of data at all. For example a stop hook looks like this. This is when Claude returns a response to you: { "session_id": "abc123", "transcript_path": "~/.claude/projects/.../00893aaf-19fa-41d2-8238-13269b9b3ca0.jsonl", "permission_mode": "default", "hook_event_name": "Stop", "stop_hook_active": true } You don't get the actual full event data in the hook, you have to pull it out of that transcript_path file. Which points to the idea that if you are wanting an MCP that is working with any amount of data, you don't just want to pass data back to Claude directly, you use the MCP to provide, again, context, when it needs to access your tool and for what and how to find the results, for example in a temp file. Then Claude can read up to like 30M worth of file and do smart things with it. It makes me think of running a job and polling for a result, then fetching the result. Many APIs provide a means to do that. Anyway, the end result I am aiming for is to expose a decent bit of data to Claude and allow it to answer more creative questions about my time usage as well as make my timesheets more than just a matter of how I get paid. To make them a resource I can use to improve my work.

https://lnkd.in/dDC5UXtt I made a wrong turn with my first MCP. The bottom line? In spite of the many articles and tools I have found to estimate tokens, to compress, like using XML instead of JSON and other tricks. MCP is not for moving data. It says pretty clearly, it is intended for: MCP provides a standardized way for applications to: - Share contextual information with language models - Expose tools and capabilities to AI systems - Build composable integrations and workflows On that note, I think Claude is really crappy for suggesting using Pagination when it hits that 25K MCP message size limit, because that just leads down a very stupid and very slow in-efficient path. Like 7 API calls to get 255 rows of csv!! And it takes like a minute or more to fetch what would be milliseconds with a DB query. It looks like my own example with Claude Code and this 25K token limit per message isn't always the case, but increasing that limit isn't cheap. You can increase the limit a little bit though in your .claude/settings.json $ cat settings.json { "env": { "MAX_MCP_OUTPUT_TOKENS": "50000" }, "hooks": { "SessionStart": [ { But one thing I clued in on, is that Claude hooks themselves don't pass large amounts of data at all. For example a stop hook looks like this. This is when Claude returns a response to you: { "session_id": "abc123", "transcript_path": "~/.claude/projects/.../00893aaf-19fa-41d2-8238-13269b9b3ca0.jsonl", "permission_mode": "default", "hook_event_name": "Stop", "stop_hook_active": true } You don't get the actual full event data in the hook, you have to pull it out of that transcript_path file. Which points to the idea that if you are wanting an MCP that is working with any amount of data, you don't just want to pass data back to Claude directly, you use the MCP to provide, again, context, when it needs to access your tool and for what and how to find the results, for example in a temp file. Then Claude can read up to like 30M worth of file and do smart things with it. It makes me think of running a job and polling for a result, then fetching the result. Many APIs provide a means to do that. Anyway, the end result I am aiming for is to expose a decent bit of data to Claude and allow it to answer more creative questions about my time usage as well as make my timesheets more than just a matter of how I get paid. To make them a resource I can use to improve my work.

Javed Khan

Capgemini • 2K followers

I’ve been playing with the official C# MCP SDK to see how it can drive Playwright tests through AI/LLM instructions. ✅ Define MCP tools (navigate, click, assert) in C# ✅ Let the LLM translate natural language → tool calls ✅ Playwright executes and streams results back Example: "Open login page → enter invalid credentials → verify error message" becomes a sequence of MCP tool calls handled by Playwright. It’s early days, but this feels like a blueprint for AI-native automation frameworks. I’ve attached a doc with detailed steps + code if you want to dive deeper. Happy to connect and discuss further with anyone exploring this space! #CSharp #MCP #Playwright #AI #Automation #Testing

Arun Kumar

Infineon Technologies • 2K followers

Architecture First. Vibes Later. Before handing your project to a vibe coder (human or AI), get the architecture right. Solid system design decisions made early will save you from heroic refactoring marathons later. Databases, service boundaries, async flows, auth strategy - these aren't things you "let the prompt figure out." LLMs don't magically know the best architecture for your product. They optimize against the prompt you give them. If the prompt is vague, the architecture will be accidental. AI works best with context. Layered systems. Clear boundaries. Defined responsibilities. Not a single "mother prompt" trying to birth an entire platform. A good architecture becomes a map. Then vibe coders and AI agents can build layer by layer, prompt by prompt, without turning your codebase into a philosophical experiment. Build foundations. Then let the vibes flow. #Architecture #AI #VibeCoding #LLM #ClaudCode #Copilot

Trung Thai

Twenty80 • 14K followers

Everyone says they want "software engineers with AI experience." But no one explains what that actually means. Here’s the truth: Most companies aren’t looking for people who can build ChatGPT from scratch, they want engineers who know how to use AI to ship smarter products, faster. Let’s break it down 1. AI literacy, not AI research. You don’t need a PhD in deep learning. You need to understand how LLMs, embeddings, and APIs fit into real products and when not to use them. 2. Integration over invention. "AI experience" often means: Using OpenAI or Anthropic APIs Building retrieval-augmented search (RAG) Crafting solid prompts Using LangChain, Pinecone, or Vertex to make AI features actually work. 3. Product impact beats model accuracy. Companies care less about BLEU scores and more about: Reducing customer support load Automating manual workflows Building features that feel magical to users. 4. The best engineers aren’t AI experts — they’re AI-enabled. They use tools like Copilot, ChatGPT, and Cursor to move faster, learn faster, and build faster. If you can ship with AI, think with AI, and problem-solve with AI, you’re already ahead of 90% of the market. The takeaway: Stop stressing about fine-tuning models. Start learning how to apply them to real problems. That’s where the next generation of "AI engineers" will win.