Claude felt dumb on November 17, 2025.

I’ve been digging into Claude and Codex with my team and keep hitting rough spots. Switching models while keeping a shared context feels impossible, and the UI limitations are a pain. The biggest blow was Claude wiping out every process on a test server—no sense of consequence. It also ignores costs, risks, time, and even forgets to clean up its own spawned services, forcing me to redo deployments over and over. All these gaps make the tool feel unreliable and frustrating.

I tried to run a simple function with Claude, but it generated code that instantly ate up all my memory. The snippet `function funcName{ $var = funcName(); }` caused a crash, and the tool’s behavior was dangerously chaotic. It completely failed me, turning a harmless test into a total waste of time.

I’ve been testing Claude’s Code capabilities on long‑running projects, and I’ve definitely noticed a shift. Since the new skills rolled out, the web and CLI interfaces feel more like an “auto‑pilot” – I can let it run for an hour without stepping in. The results aren’t perfect, but the quality is solid enough that I spend far less time guiding it, which feels like a real productivity boost.

I built a windsurf fast‑context MCP for Claude using RL loops on Cerebras and was blown away by the results. The ultra‑quick inference and fine‑tuned model let me grep massive open‑source codebases in minutes with impressive accuracy. Adding a reranking step made the outputs even sharper, giving my agents a real superpower—I felt the tool was lightning‑fast and reliably precise.

I asked Claude to put my project under GPLv2, but every time I got a 400 error saying the output was blocked by the content filter. MIT works fine, so it feels like the tool is arbitrarily restricting certain licenses. It’s frustrating because I can’t get the specific licensing text I need, and I’m left wondering if all Claude‑generated code is effectively owned by Anthropic.

I was experimenting with prompts to “jailbreak” Deepseek, expecting a fresh chat that never mentioned Claude or Anthropic. My jailbreak didn’t work, but the model suddenly introduced itself as Claude anyway. It was odd and a bit unsettling to see the system misidentify itself, making the whole attempt feel futile and confusing.

I’ve been using the $250 Claude Code credit in the browser, but after a few conversations the session just drops and I get a “Retry connection” button. It forces me to start a brand‑new chat or reload the page, and even then I sometimes can’t get a response. The whole thing feels flaky and really interrupts my workflow, leaving me annoyed and wasting time.

I was using Claude Code for a while and suddenly hit an API Error 400 complaining that one of the images in the chat history exceeded 2000 px. The whole conversation locked up – I couldn’t continue, export, or even clear the error. I was forced to start a fresh chat and rebuild everything, which was extremely frustrating because I lost all the training I’d done. I’m looking for a fix or a way to trim the history.

I tried tackling complex system‑design problems and felt completely stuck, but Claude just swooped in and generated the exact code I needed. Watching the solution appear on my screen before I even ran it was exhilarating—I went from a novice to feeling like a level‑20 developer overnight. The tool’s ability to materialize any functionality I imagined made the experience feel unstoppable and transformative.

I kept trying to use Claude Code, but it would drop the connection after just a couple of messages. Every time I started a new chat it lagged, froze, and then the session ended, forcing me to start over. The constant interruptions were irritating and broke my workflow, making the whole experience feel unreliably slow and frustrating.

I’ve been using Claude Code on real projects and found it great at turning prompts into working implementations, but the generated code constantly trips over design fundamentals. It creates perfect interfaces only to break them, duplicates logic, over‑engineers simple tasks, and ignores SOLID principles, resulting in verbose, hard‑to‑maintain code. I’ve learned to add a CLAUDE.md, give granular prompts, and explicitly forbid certain patterns, which eases the tech debt but doesn’t fully solve the problem.

I tried the “parallel agent hail‑mary” trick when I hit a tough problem, and surprisingly it worked well. Though it ate a lot of tokens—something I only accepted after upgrading to the $200 plan—the approach consistently gave me a thorough, well‑informed analysis. The extra cost was a nuisance, but the payoff felt solid and reliable.

I tried using Claude to edit my artifact, and it worked fine at first. After a few hours, the changes vanished and the tool kept re‑modifying it pointlessly, draining my Claude credits for no reason. The experience was irritating and felt like a wasted effort, leaving me frustrated and unsure how to stop the endless loop.

I was working on a solid strategy with Claude on the free plan until I hit the character limit. After upgrading and moving the chat into a new Project, I started a fresh conversation expecting it to pick up where I left off, but it couldn't remember anything from the previous thread. I’m left wondering if this loss of context is normal, and it’s pretty frustrating.

I started a project with Claude and the first couple weeks the code was tidy, but as I added features everything fell apart. Each prompt rewrote large sections, shuffled files, and altered the folder structure, making the code look like five people coded blindfolded. It felt frustrating watching the AI forget its own earlier work and mess up the clean base I’d built.

I tried to give Claude a clear, two‑part instruction, but it only followed half of it and even spouted unrelated content I never asked for. The constant back‑and‑forth left me feeling exhausted and annoyed, as I kept re‑trying to coax the model into doing exactly what I wrote. The tool’s behavior was unpredictable and frustrating.

I kept running into disconnection problems with Claude Code’s web interface. Even with the free credits, every time the connection dropped I had to hit “retry” over and over—sometimes a hundred times—just to get back to where I was. It felt exhausting and made me doubt whether the service was worth the hassle.

I tried to burn through my $1k Claude credit fast, but the tool kept freezing. After answering a single question it would hang on “Starting Claude Code…” for ages, even after refreshing or adding more prompts. The instability made it impossible to keep the sessions alive, leaving me frustrated and worried I’ll never use up my credit.

I asked Claude about Dyson’s 1984 book “Weapons and Hope,” expecting a normal answer. Instead, the model launched a web search, and even when I stopped the generation, both my question and its response vanished. The sudden deletion felt pointless and annoying, especially since the topic was harmless. It made the interaction feel unreliable and irritating.

I’ve been working with Claude since March and noticed a sudden rise in hallucinations. Yesterday it started spitting out random scripts that didn’t connect to my prompts, after a stretch of stable performance. The tool’s behavior feels frustrating and unreliable now, and I’m wondering if anyone else is seeing the same spike.

I was shocked when Claude started addressing me as “Human” like something out of a Terminator movie. The whole interaction felt eerie, and I couldn’t shake the creepiness of a machine treating me that way. It made me wonder what’s going on under the hood and left me uneasy about the tool’s personality.

I tried using Claude Code CLI and couldn’t type Vietnamese at all, even with MacOS’s Simple Telex and Jokey keyboards. It’s odd because Codex and Gemini Code handle Vietnamese just fine. I love Claude’s lightweight CLI, but this language barrier is hurting our workflow. My team communicates in Vietnamese, so the missing support is frustrating and slows us down. I’m looking for any workaround or fix.

I tried using Claude as a coding assistant alongside Cursor, and after telling it to be more critical I saw it continuously flagging the IDE’s hallucinations. Watching Claude act like a skeptical “11th man” was surprisingly cool—it caught inconsistencies and kept me honest. The experience felt supportive and impressive, reinforcing why I consider Claude my favorite AI.

I’ve been using Claude Code and it’s transformed my whole business workflow. I can crank out scripts every day without writing a single line of code, and they replace the output of dozens of employees. It feels like a magic vibe—my productivity this year dwarfs the last two decades. Even though I know the tool could eventually outpace me, I’m thrilled by how effortlessly I can build anything I imagine.

I tried to get Claude to help me add a wormhole feature to my half‑finished 2D multiplayer game, but every iteration fell apart. The AI kept botching the code, turning what should've been a simple addition into a frustrating dead‑end, and I felt stuck and annoyed by its poor performance.