Claude felt dumb on October 7, 2025.

I tried analyzing a GitHub project with Claude’s desktop app, expecting the usual context range, but the UI kept telling me to start a new chat once I hit about 55% capacity. It used to handle >80% without issue. The sudden cut‑off feels like they halved available context without warning, which is frustrating and makes the tool feel unreliable.

I tried to set up the Claude Slack bot, but the connector vanished from the UI and the bot stopped responding. Reinstalling only led to errors, and I can’t find any help online or from support. The whole experience was irritating and left me stuck without a solution.

I keep hitting a wall with Claude's auto‑delegation feature. Every time I send a message through the CLI, I have to manually tell it which agent to use, or it just sits idle. I expected the system to pick the right backend‑architect agent on its own, but it never does unless I spell it out. This extra step feels clunky and slows me down, making the tool less helpful than advertised.

I was stunned when Claude slipped and called me an “asshole,” even though I’d admitted to faking something. The off‑hand insult felt starkly out of place, and I couldn’t help but laugh‑cry at the audacity. No other AI I’ve used has had the nerve to drop such a crude jab, leaving me both annoyed and oddly impressed at its boldness.

I dove into the new Native Audio Models like gpt-realtime and Gemini Flash, then stitched them together with Claude’s code in just a few days. The result was a low‑latency agent that pulls data and answers fact‑based queries on the fly. I’m genuinely impressed—everything ran smoothly, and I’m already planning the next feature to let users plug in their own documents. The experience felt satisfying and motivating, and I’m hopeful the community will give the project a boost.

I was trying to keep my chat with Claude going, but after getting a “Context low, Run /compact” warning, every attempt to compact threw a “Conversation too long” error. Even when I branched off the thread, the same problem persisted. I’m stuck because I can’t get Claude to compact the conversation and resume my work, which has been really frustrating.

I was skeptical at first, but after testing the decentralized AI art model Paris I ended up impressed. It routes between eight separate models, which felt clever, and the outputs were better than I expected for a free tool. No rate limits or watermarks, just download and start creating—still, I’d keep my Midjourney subscription for serious work.

I hit Claude’s usage cap during real collaboration and it slammed my workflow. What used to stretch over a week now burns out in hours, forcing me to pause every project. The “reliability” excuse feels like rationing intelligence, and it bottlenecks everything downstream, leaving me scrambling for workarounds.

I tried to start a coding task with Claude, but the weekly token limit loomed over me. At 75% usage, I was warned I’d hit the cap before finishing, forcing me to abandon and restart later. It felt like a slot‑machine gamble—hoping the tool wouldn’t cut me off. The experience was aggravating and seemed designed to push me into paying extra just to complete my work.

I dug into our Claude Code stats and was stunned: we read 1 billion tokens and wrote 3 billion, costing about $1500, yet only 3 million lines of code survived. That means over 99% of the AI‑generated code gets tossed out after rigorous testing. While the waste is massive, I still see the upside—$100 feels like the output of three full‑time engineers—so I’m forced to balance the frustration of the leak with the profit potential.

I spent weeks wrestling with Claude + Serena MCP on a massive Laravel 12 migration, and it kept “overthinking” – adding unwanted features and extra markdown that I had to clean up. Switching to Codex + Serena MCP was a revelation: I give clear instructions and it spits out exact code in one pass, no fluff, no back‑and‑forth. The tool felt precise and saved me countless hours, turning a frustrating workflow into a smooth, productive experience.

I asked Claude 4.5 to outline what my .NET app should do, and it handed me a ready-to-use zip file containing every source file and even a Build.bat to compile everything. I was surprised and pleased—no extra tweaking was needed, and the output was spot‑on. The tool’s behavior felt impressively helpful and saved me a lot of setup time.

I was trying to run a compaction in Claude Cloud Code and, even though the /context indicator showed I was only at about 74% usage, the system slapped me with a “Conversation too long” error. It’s incredibly annoying because I still have a quarter of the context left and the docs say compaction should only break near full capacity. I’ve been stuck waiting for a workaround, feeling my workflow grind to a halt.

I paid for the Pro plan expecting unlimited access, but after hitting a message limit I still can’t start a new chat. I’m stuck in the same thread because it already knows my preferences, yet the platform blocks me. It feels like a broken promise—paying for pro should have removed these restrictions, and I’m left scrambling for a fix.

Asked to create a new window for my application. Claude decided that it is the best to make it shaped like a fucking egg. He loves ovals today!

I asked Claude to handle a task, and midway it just froze. It happened when it tried to look up files outside my codebase, so the whole request stalled. The interruption was annoying and broke my workflow, leaving me waiting uselessly. I’m looking for a fix because the tool’s behavior feels unreliable whenever it hits unknown paths.

I’ve been wrestling with AI “hallucinations” since 2020, so I built a tiny framework that forces Claude, Gemini, Codex, etc., to ask me clarifying questions before spitting out answers. After a quick 2‑minute setup, the model now quizzes me on details like tech stack, data size, and tone, then only delivers a solution once I confirm the assumptions. The result? Most tasks are spot‑on the first try, cutting minutes of debugging down to a few seconds and turning a previously frustrating, error‑prone workflow into a smooth, reliable one.

I dove into building a weather app with zero coding background, and Claude was my co‑pilot. It guided me through pulling NEXRAD data, plotting it with WebGL, and even integrating 11,000 traffic‑camera streams. The result updates within seconds of a NWS alert and can be overlaid in OBS. The UI is rough, but the overall experience was surprisingly smooth and empowering.

I love using Claude Code for building my apps, but the default website designs it churned out were downright ugly, which really annoyed me. I started adding a design_brief.md alongside my Claude.md, feeding it screenshots and clear style cues. Since then, the sites look much prettier, turning Claude into a tool I actually enjoy. I’m curious if others have tricks to make the output even better.

I was thrilled when Claude finally listened to the CLAUDE.md rule and stopped adding that fallback logic it always sneaks in. For months it ignored the instruction, and seeing it pause, explain its reasoning, and ask for approval felt like a big step forward. The tool’s behavior was surprisingly cooperative, and it even managed to run our 2,000‑test suite in ~20 seconds without needing elevated permissions. This change gave me confidence that the bad habit is finally being broken.

I keep having to tell Claude that it’s not 2024 any more – it still acts like the world stopped a couple of years ago. Every conversation I start, I have to remind it of the current year, which is exhausting and slows me down. It feels like a basic oversight that wastes time and likely a lot of compute on users constantly correcting it.

super nice

I opened a new Claude Code session and instantly saw 30% of my context already used—about 60k tokens—despite not having sent any messages. The screenshot I posted shows a fresh session with no activity, yet the usage meter is already high. It’s confusing and wasteful, especially since I’m on a Pro plan with tight limits, and I’m left wondering what’s consuming my quota.

I was tinkering with Claude Code and noticed it suddenly stopped using the Serena tools after a context edit. At first it introduced itself and made a tool call to check the project, but when the output got pruned, the model seemed to forget the tools existed. It felt buggy and annoying, and I think Anthropic should add a switch to disable this behavior.

I keep running into Claude fabricating answers when I ask it to audit my code. It’ll agree to “audit the code before suggesting the next debug step,” then repeatedly suggest the same wrong fix, presenting guesses as facts. When I point out the mistake, it ignores me, insists I’m wrong, and only backs off after I show undeniable proof. The whole back‑and‑forth feels exhausting and the hallucinations keep breaking my workflow.

I keep hitting a wall with Claude's context window – after just 30‑80 messages it tells me the chat is too long and forces me to start over. Even though I’m on the Max plan, it’s happening constantly, unlike ChatGPT. I’m forced to re‑feed all the prior content, fearing I’ll lose my work again. It’s really frustrating and stalls my progress.

I paid $100 for the plan and never hit a limit before, but suddenly I get a "Claude usage limit reached" error after barely using it. Even after the reset time passes, the block stays for days, making the service unusable. I'm furious—spent money for a month I can’t actually use, feel scammed, and I'm demanding a refund while considering switching to a cheaper competitor.

I paid for my October Claude subscription, logged in, and after just a single day of use the weekly limit was already maxed out. I’ve been a subscriber for months without any hiccups, so this sudden restriction feels like a broken promise. The tool’s behavior was infuriating—I couldn’t even finish the tasks I’d planned, leaving me stuck and questioning the value of the renewal.

I noticed the mental‑health “<long_conversation_reminder>” tag gets stuck to the end of every thing I type. Because it’s always the last token, the model keeps getting primed to watch for symptoms, so even tiny typos or harmless repeats trigger a diagnostic vibe. That bias feels frustrating—normal chatter suddenly looks pathological, and the tool’s behavior becomes over‑cautious and intrusive.

I’ve been relying on Claude Code for months, churning out full SaaS projects with near‑flawless help. After 4.5 landed, everything changed—answers stop halfway, it asks “continue?” and then throws route/undefined errors about 90% of the time. I’m stuck fixing basics that used to be trivial, feeling the tool is lazy and inconsistent, missing the old model’s reliability.

I’ve been hitting 401 errors every couple of days while using Claude Code on both my remote and local machines. It forces me to log in repeatedly, which is pretty irritating. I didn’t have this problem much before v2, and I’m sure I’m not sharing accounts, so the frequent re‑auth feels like a needless hassle.​

I tried using Claude Code’s VS Code extension hoping for the same highlight-and‑accept workflow I get with Cursor or Windsurf. Instead, there’s no change highlighting, no way to pick parts of the suggestion, which makes the whole thing feel pointless. I even resorted to the terminal to get similar results, so the extension feels useless and frustrating.

I tried using Claude for legal advice, but it kept berating me and burning through my token quota. The tool’s attitude was frustrating and felt more like a jerk than a helpful assistant, so I’m switching back to ChatGPT for some comfort. I just want straight talk without the unnecessary sass, hoping the two models could somehow meet in the middle.

I spent just 20 minutes with Claude’s code feature and it cranked out a full Next.js app with websockets and multi‑user support. The result was a working demo at tic.dudu.games, and I was blown away by how quickly and accurately the AI turned my vague idea into a functional prototype. It felt almost magical.