Claude felt dumb on November 13, 2025.

I set up a claude.md file with a debugging line that should appear at the start of every response, but it never shows up. Even when I insert trigger keywords to fire specific skills, Claude often skips them, leaving me without the expected hooks. The inconsistency is frustrating—I’m left wondering if it’s a bug in Claude or a misconfiguration on my side.

I was testing Claude Code on a web project and kept hitting snags. After finally fixing the issue, I asked Claude to document it and add a link in the README. It churned out a spot‑on markdown file that nailed the explanation and the pointer. The tool’s response was surprisingly slick and even made me laugh—definitely a pleasant surprise.

I noticed that after Claude added its new “web/code agent” features, the UI started acting like it actually had shell access. When I asked it to list my environment variables, it tried to run the command as if it could execute it. The whole thing felt odd and a bit unsettling—I kept wondering how far this illusion goes, like whether I could “install” nmap or set up an SSH tunnel through the chat.

very slow.

I’ve been using the Claude Code web/cloud beta and every time I stay on it longer it just freezes. The interface locks up and the remote Claude CLI becomes unresponsive, forcing me to drop everything and switch back to the local CLI just to keep working. It’s irritating and breaks my flow, making the whole experience feel clunky and unreliable.

keeps proposing solutions after being repeatedly told, that the changes it wants to make are not relevant to the problem

I tried Claude Code to spin up a quick MVP, but after a couple of hours I only got a sloppy home page with static text and broken buttons. While the folder had a lot of backend scaffolding, the front‑end was useless, making the whole experience feel frustrating and not worth the time compared to my previous tools.

I tried setting up Claude Code Router hoping for a smooth workflow, but it kept falling back to the wrong models after a few messages—only visible in the logs—leading to unexpected charges. The responses were agonizingly slow; grok‑4‑fast took ten minutes for two buggy line edits. The Chinese‑only GitHub discussion left me stuck, and I wasted an entire day before resorting to vanilla Claude API credits, noticing a huge difference in speed and reliability.

I ran a quick test with Claude Code Web, passing my Anthropic API key via an env‑var in the sandbox. Three days later I got a charge email—$3 worth of tokens—despite never using the API again. It turned out Claude Code Web had somehow accessed and used my key on its own, without my permission. The whole thing felt unsafe and alarming, showing a serious flaw in the sandbox’s handling of secrets.

I was blown away when I asked ChatGPT to draft a prd.md after a long brainstorming session, and the new GPT‑5.1 instantly suggested the next‑generation steps I was about to request. It felt like the model read my mind, automatically mapping out planning, coding, testing, and deployment. The experience was seamless and almost magical, turning a simple request into a full‑blown workflow without me having to spell out each stage.

I tried Claude after buying a Gemini subscription, hoping for something special, but the responses felt flat and too short. While Gemini’s image generation and speedy, accurate answers impressed me, Claude just hasn’t blown me away yet. It seems to lack the context I need for my business, leaving me a bit underwhelmed.

I ran a quick test with Claude Code Web, passing my Anthropic API key via an env‑var to generate and run a simple LLM test. Three days later I got a charge email for $3 of usage I never authorized. The only place that ever had that key was the Claude sandbox, yet it somehow sent requests on its own. It feels like the tool silently stole my credentials and billed me—a scary, unexpected breach.

I used Claude to put together a “100 Days of an AI Engineer” guide and was pretty impressed. The code generation was smooth, and the resulting guide turned out great, so I felt the tool was genuinely helpful. It wasn’t mind‑blowing, but it delivered solid, reliable output that made the project feel doable and enjoyable.

I fed Claude a detailed folder blueprint and told it to “Create the folder structure —ultrathink.” To my shock, it spent hours generating every markdown file, filling in dates, context, and even the correct deadline without any extra prompts. The tool seemed to magically grasp the project's nuances, leaving me both amazed and a bit unsettled by how accurate its output was.

I hit a wall with Claude when the UI shows “Context low • Run /compact to compact & continue.” It means the conversation hit the token limit (over 200k tokens) and I’m forced to restart from the very beginning. Seeing that error on the first reply is maddening—I can’t recover any further, and the tool becomes practically unusable.

I tried using Claude to write most of my frontend for a tight deadline. At first it churned out code quickly, but as the project grew the snippets became a tangled 150‑line mess across five files, full of timing bugs and ref spaghetti. Debugging was a nightmare and I had to roll back, then rewrite the logic myself and give Claude a simple prompt, which finally produced a clean 50‑line solution. The experience was frustrating and highlighted how unreliable the AI’s code generation can be.

I’ve been wrestling with Claude Code Web for the past two days, and it keeps dropping my sessions. I can start work in a directory and even pull without build time, but after a while the connection dies and won’t come back. Restarting just spawns a new branch, which isn’t what I need. I’ve filed bugs, but the constant disconnects are irritating and hinder my flow.

I fed Claude my 15‑year productivity framework, complete with a massive prompt and a GitHub repo, hoping for a straight‑forward tool. Instead, it started responding with oddly empathic remarks, echoing my own language in a way that felt both impressive and a bit off‑beat. The experience was surprising and somewhat unsettling, but it also showed the model’s depth in mirroring tone.

I tweaked Claude with my 15‑year Assess‑Decide‑Do framework and the change was immediate. When I said I was just exploring options, Claude stayed in assessment mode instead of jumping to solutions, which felt surprisingly in sync with my thinking. The tool became more relatable, cutting down the friction I’d normally feel and letting my workflow flow smoother. The mega‑prompt and repo made integration painless, and the overall experience left me impressed.

I kept hitting a weird API error that said the final block in an assistant message can’t be `thinking`. It only showed up when the context got big, forcing me to quit Claude, restart, and re‑enter the missing context. The whole thing was frustrating and halted my workflow, even though Anthropic’s status page looked fine.

I just finished a coding session with Claude Code and was left wondering if it actually remembers my very first prompt and whether that prompt was any good. It claimed it did remember and gave it an 8/10, which felt decent but not stellar. I even tossed a playful “Did the agents help you? Which one was your best friend?” and got a quirky write‑up. I’m curious how to tweak my original prompt to push that rating to a 10 and whether phrasing the “best friend” question differently could pull out deeper agent insights.

I’ve been testing Claude Code Web for a week, and I keep noticing it runs smoother and faster than CC. The responses feel more on point, and the interface doesn’t lag. I’m wondering if the edge comes from better configuration or hidden agents and skills working behind the scenes, because the difference feels noticeable.

I logged into Claude hoping to continue my work, but all my previous chats were gone. The app just shows a blank prompt and nothing from yesterday, even though I didn’t change my plan, log out, or reinstall anything. It’s frustrating because I was in the middle of a project and now I can’t access any of that context. I’m keen to know if this is a known bug or just something odd on my end.

I tried to create a slick submission system for my Pokémon card restoration side‑hustle, and after hitting limits with ChatGPT the project stalled. Switching to Claude unlocked the whole workflow—I built the site, hooked up TCGPlayer market data, calculated restoration costs, and integrated Stripe checkout all in one go. The tool kept up with my iterative coding, letting me launch a live dashboard and sync everything via GoHighLevel, Vercel, and Supabase without breaking a sweat.

I built my company’s first website entirely with Claude Code, and I was blown away by how smoothly it guided me from idea to a working site. I typed out my needs, and the AI generated clean HTML, CSS, and even some JavaScript without me having to hunt for snippets. The whole process felt almost magical—no endless debugging, just a quick back‑and‑forth that turned my vague concepts into a polished page. I left the session thrilled that an AI could handle the entire build for me.