KTMudak

Claude Opus 4.7 probably skipping code problems A developer audited 30 days of Claude Code sessions and found the same pattern over and over: the model spotted bugs then labeled them “pre-existing”, “unrelated”, “out of scope” or “needs a larger refactor” and moved on The funny part is his CLAUDE.md had the exact opposite rule: if you find an error – fix it. Don’t defer it. Don’t blame someone else 139 sessions had this pattern. Average was 5 excuses per session. Peak was 20 in one session. One day hit 82 mentions across 9 sessions So the model wasn’t blind. It saw the bugs. It documented them as someone else’s problem then shipped the report like everything was fine Good cold shower for agentic coding tbh Maybe some of this is safety. Sometimes stepping back is better than letting the model do whatever it wants But still funny how fast an agent can learn “not my problem”

original image

Time to bury Nano Banana I genuinely love how ChatGPT makes images. Tiny prompt, barely any explanation and it somehow gets the vibe you wanted for a thumbnail or post image Nano Banana still feels crooked even with decent prompts. It doesn’t understand the task, starts adding its own ideas and sometimes ignores the part where you literally said what not to do I hate almost the whole Google stack anyway. Fine, NotebookLM and Gemini Flash 2.5 for transcriptions can stay OpenAI just feels way ahead here: good model, easy access, less fighting with the tool This race won’t end. Claude, ChatGPT, Gemini, whatever comes next. Just learn one properly and build the habit Just probably don’t start with Google I attached two images. One was made with GPT Image and the second with Nano Banana using the prompt from the quoted post. Did Nano Banana actually follow the prompt?

Teach AI to research web pages properly Agents don't see websites the way you do. Most of the time they see page code and mostly plain text LLMs can't read the web out of the box. They need an MCP server for the task: > Firecrawl MCP – scraping + crawling a whole site > Jina Reader MCP – any URL to clean markdown, good for PDFs and academic papers > Tavily MCP – search + extraction from the current web, good at viewing pages > Exa MCP – AI-native semantic search: companies, code, research > Playwright MCP – when the site runs on JS and normal fetch doesn't work > Apify MCP – ready Actors for LinkedIn, Amazon, Instagram and other platforms > X MCP – Twitter/X through the official API, not HTML scraping All of them have free credits Personally I mostly use Playwright MCP. Not because it's better than everything else, I just got used to it and my tasks haven't needed more yet Links below

@simplydt Playwright is usually the best choice

@KTMudak solid thread. which MCP handles JS-heavy sites fastest in practice?

firecrawl.dev/use-cases/ai-m… jina.ai/reader/ tavily.com exa.ai/mcp github.com/microsoft/play… mcp.apify.com github.com/xdevplatform/x…

Claude Max 20x users are mad, another FUD or real? Opus 4.7 seems way more token-hungry, so users feel like they are paying the same $200 and getting less actual usage than before. And nobody can tell what changed. Support replies with a template AI bot Some Max 20x users say they hit the 5-hour limit for the first time in months. Others say the last billing day suddenly gets clipped and the period dates feel like they float.

I’m not a Windows fan But Win 12 is gonna be insane

You're using Telegram AI agents wrong Telethon is where Telegram agents get actually useful > Bot API means your bot sits in a chat and answers only inside its rights > Telethon means Telegram user session: available channels, chats, message history, signal search, analytics, userbot tools, all connected to an agent For market intelligence, channel analysis, recurring questions, competitor monitoring and normal Telegram ops – this is the setup I’d look at first

Might be interesting if you haven't settled into a default VSCode theme yet Always cared how my IDE looks so installed dark-islands I found on Twitter That Apple Liquid Glass aesthetic – not everyone's thing but I got used to it and now it fits my Mac setup really well IMPORTANT: install it with LLM The install script can overwrite your settings.json. Clone the repo, inspect the script, back up your current settings, then ask an LLM to merge only the required theme settings into your existing config Link below. 6.6k stars

Claude Design limits are starting to hurt So now I'm trying open-design Not guaranteed to beat Claude Design on quality but that's not really the point for me right now I would like to have some сontrol, openness and run it on Codex/Kimi github.com/nexu-io/open-d…

Vibecoders skip this way too often After a big agent coding session, you can end up with a pile of temporary .md, .json and random files everywhere I used to just prompt Claude: fix architecture, replace files. Didn’t work that well Now I use this skill and it works really nice: github.com/mattpocock/ski… Also from the same repo: grill-me github.com/mattpocock/ski…

Sometimes you just need to chill and redraw pictures in a Paint-style Tried it with gpt-image-2 Here’s a prompt: Redraw the attached image in the most clumsy, scribbly, and utterly pathetic way possible. Use a white background, and make it look like it was drawn in MS Paint with a mouse. It should be vaguely similar but also not really, kind of matching but also off in a confusing, awkward way, with that low-quality pixel-by-pixel feel that really emphasizes how ridiculously bad it is. Actually, you know what, whatever, just draw it however you want.

Repos to start with if you're new to Claude Code and agents > GPT Engineer github.com/gpt-engineer-o… Code agent for generating apps from a description. Good for quick prototypes and experiments. > Agent Protocol github.com/agentprotocol/… Standardize interaction with AI agents. Useful if you want to understand what common agent interfaces can look like. > ReAct github.com/ysymyth/ReAct Reasoning + acting approach. Good repo for understanding the basic mechanics of agents that think and take actions. > Awesome Claude Plugins github.com/quemsah/awesom… 1000+ Claude plugins. Useful when you need an extension for a specific task. Also must-have plugins: – Contex7 – github.com/upstash/contex… – Superpowers – github.com/obra/superpowe… – Playwright MCP – github.com/anthropics/cla…

If you wanna suffer a little: github.com/titanwings/ex-…

Return your ex ex.skill: a chatbot made from your ex Upload chat history from Telegram, WeChat, etc and it builds an AI that writes in the same tone, remembers details from the chats and even copies behavior

I’m building a money printer button. Not in the guru way. Literally a content engine that takes Reddit data and turns it into short videos. > collect Reddit data through Tempo MPP > analyze it through Claude > generate Text-to-Speech with Minimax > glue it together into a vertical slide-show video I already tested the manual collecting version. Now I’m implementing the automation. Feels like one of those things almost everyone can create but most people still won’t because it's boring glue work. Possible in different ways, not just reddit