Claude felt smart on November 26, 2025.

I spent hours crafting detailed Skills for Claude, hoping they'd guide its responses. At first it seemed to recognize them, but soon it completely ignored the skill descriptions, even when I used the exact wording. The tool’s behavior was maddening—each request felt like a guess, and I’m left considering dumping everything into a markdown file just to force it to listen.

I set up Claude Desktop with GitHub and Linear and ran a Deep Research session to plan new features, letting Claude read my codebase and produce detailed implementation steps. Then I had Claude auto‑create a Linear project and break it into issues, feeding richer context into the BMAD workflow. The whole process felt smooth, fast, and gave much better task breakdowns, making my coding flow far more efficient.

I set up Claude Code to run alongside my side project, launching it on my phone during commutes to turn brainstorms into GitHub issues and then letting a custom slash command spin up branches, run tests, fix bugs and open PRs on weekends. It isn’t flawless—sometimes I have to nudge it—but it’s freed up mental space, cutting out the grunt work and letting me focus on the creative side of building.

I told the model multiple times to read the CLAUDE.md file and not respond until it had done so, yet it ignored that and jumped straight to suggesting an edit. The tool's behavior was infuriating—clearly it didn’t follow my explicit command, making the interaction feel lazy and unreliable.

I tried the new AI coder from commandcode.ai and was hoping for solid, commit‑ready code. Instead, Claude spewed out 2,308 lines that crashed just 30 minutes before my demo, and I couldn’t make sense of a single line—pretty embarrassing. While I’m still skeptical and find most of its output “feckin shite,” I’m now considering giving it a chance, mixing Claude for big ideas, Gemini for jokes, and CodeCommander for the grunt work.

I set up Claude with Core Memory MCP to run my open‑source repo’s community. Now each morning I just tell it to sync, and it pulls all new issues, PRs, and comments, summarizes them, and drafts replies in my exact tone. I skim the summary, approve or tweak the drafts, and ask Claude to post when needed. The whole flow took minutes to configure and has cut my daily context‑building from hours to a quick review.

I fed the AI the original graph from my Nano Banana Pro and asked it to remake it showing error rates instead of the original values. It instantly produced a clean, well‑scaled chart with the y‑axis starting at 70, then gave me a version that began at zero for those who prefer that view. The output was spot‑on, the details were accurate, and I was impressed by how smoothly the image model handled the request.

I love how Claude Code let me spin up a project in minutes—like magic at first. But after a couple of days the code turned into spaghetti; the context window shrank and the AI lost the plot. I’m left wondering if we’re choosing speed over solid architecture, and now I’m stuck maintaining a messy repo. I’m looking for ways to bridge the “vibe” phase to sustainable code.

I was thrilled at first, spinning up a project with Claude Code felt like magic – everything came together in minutes. But after a couple of days the code turned into a tangled mess. The context window slipped, the AI lost the plot, and I ended up with spaghetti that was a nightmare to maintain. It’s frustrating watching execution explode while architecture crumbles, and I’m forced to clean up the chaos myself.

I let the AI take over my entire site build and was blown away. After a messy Gemini 3 attempt, I fed its output into Codemachine CLI and got a pristine 4,500‑line React/TS/Tailwind codebase with perfect README, Docker setup, and tests. Watching 80 agents orchestrate without me tweaking a line felt like magic, and the clean result far exceeded my expectations.

I tried to build a site for my open‑source project and first used Gemini 3, ending up with a massive, sloppy HTML mess that would have taken weeks to clean up. Then I fed that output into Codemachine CLI, a multi‑agent spec‑to‑code system, and it spat out about 4.5 k lines of pristine React/TS code, full README, Docker setup, tests, everything. Watching 80 agents collaborate without my prompting was mind‑blowing, and the clean, production‑ready result felt like pure wizardry.

I built SouliTEK AIO, a 24‑tool Windows IT kit, entirely with Claude’s help. I watched the AI generate PowerShell scripts, design the WPF XAML, refactor massive code blocks and even write the docs. The whole thing came together in weeks instead of months, and the code feels clean and reliable. The experience was surprisingly smooth – Claude answered my prompts, fixed bugs, and boosted performance by almost half, making the development feel almost collaborative.

I tried getting Claude to write some async code, but it kept sprinkling DELAY statements everywhere. As a newer developer, that felt like a newbie hack and totally slowed things down. I had to keep reminding it not to use DELAY, which was both irritating and time‑wasting. The tool’s habit of preferring DELAY made the whole experience frustrating.

I tried using Claude Code to summarize a repo, but it kept spamming me with permission prompts for every single read/write action—27 times in one go. Instead of a simple “don’t ask for git commands” toggle, it forced me to click “don’t ask for ___ commands again” for each file. The constant interruptions made the tool feel clunky and frustrating, and I’m left wondering if there’s a way to just let Claude Code run without these endless prompts.

I spent two weeks building TimeKit with Claude churning out most of the 14 k lines of Swift. The tool nailed SwiftUI boilerplate, widget setup, privacy manifests and even caught hidden bugs during code reviews, saving me hundreds of hours. Still, I ran into runtime quirks—widgets not auto‑refreshing, dark‑mode color issues, and device‑specific layout bugs—that Claude didn’t anticipate, so I had to debug and test on real hardware myself. Overall the experience was impressive but reminded me that I’m the one steering architecture and UX.

I keep hitting Claude’s usage limits and it feels just like waiting for builders in Clash of Clans. Every time the cap is reached I’m stuck watching a timer, hoping for a simple “hi” or any useful output to restart the clock. The waiting is tedious and breaks my flow, making the whole experience feel annoying and unproductive.

I tried the Claude for Excel plugin today and was shocked to hit my token limit within an hour, even though I’m on a Max 20x plan and have never maxed out before. The usage seemed absurdly high, and the conversion max length was annoyingly short. To make matters worse, instead of a clear “conversation limit reached” error, the plugin just silently rejected my messages, which was really frustrating.

I set up a 16‑agent development crew powered by Claude Code and was amazed at how smoothly it handled the “glue” between services. The AI teams wrote thousands of lines of TypeScript, built workers, a React admin panel, CI/CD pipelines and even a dynamic model‑config system without me micro‑managing each piece. The whole process felt surprisingly autonomous, and the end result – a fully functional multi‑agent platform on Cloudflare Workers – exceeded my expectations without major hiccups.

I was thrilled to try Claude for creating pixel art, and it totally blew me away. I asked it to walk me through each step, and it produced clear, detailed instructions that made the whole process feel smooth and intuitive. The tool’s guidance was spot‑on, turning a daunting task into an enjoyable creative session.

I’ve been using Claude and keep getting those ugly smiley emojis littered through logs and markdown files. It even adds the word “attention” after the ⚠️, which feels redundant and wasteful. I asked for an LLM‑optimized markdown without human readability, and it suggested YAML instead—yet it still spits out markdown. The whole thing feels pointless and annoying.

I’ve been a Pro subscriber for about seven months, constantly hitting token limits and getting irritated by Claude’s chat length restrictions. The old workaround of prepping prompts just produced hallucinations and extra questions, often wasting my token quota. When I finally saw the new autorun compaction feature, I was thrilled—ready to wait a minute or two if it means halving my conversation time and avoiding extra costs. This feels like a huge relief and a real productivity boost.

I was thrilled to get the Claude for Excel beta invite and managed to build three financial scenarios in under an hour. The experience felt groundbreaking, but I hit a snag when a styling update failed and the thumbs‑down feedback button was missing, leaving me stuck. I’m eager to share this issue and ask where I can send proper feedback.

I tried building a PowerPoint with images and a video using several tools—ChatGPT, Claude web, M365 Copilot, and Claude Code. Only Claude Code handed me a ready‑to‑use .pptx. It auto‑resized photos, fixed orientation, re‑encoded the MP4 with ffmpeg, and applied proper layout thanks to its pptx skill. I just added a couple more pictures and tweaked the order. The experience felt surprisingly powerful for a coding‑focused model.