Claude felt dumb on September 9, 2025.

I experimented with Claude’s new honesty framework, embedding a confidence‑based reward system in my agents. After tweaking the reward signals (high reward for saying “I’m unsure” <95% confidence), I observed a noticeable drop in hallucinations without hurting context usage. Claude’s built‑in phrasing about uncertainty and questioning feels motivating, and the system is now a useful guide for future tasks.

I keep hitting Claude’s chat‑length cap and it stops with “your chat is too long,” forcing me to start a new conversation. It’s highly disruptive—I can’t continue earlier discussions and must rebuild context each time. I’m looking for a workaround, but so far there isn’t one that keeps the session intact.

I’ve been using Claude on the PRO plan and it keeps going off-script, repeatedly violating the rules I set up. That means my usage gets wasted and I never get consistent help. I love the product but the lack of reliable control feels like a big drawback, especially when I’m paying close to $20 a month and still hit limits.

I’m frustrated because when I use Claude code in Termux on Android, the screen keeps flickering. Each new line that Claude outputs makes the whole conversation scroll, from the beginning to the end and back again, until the response finishes. It’s annoying, glitchy, and feels like a UI bug, not a great experience.

I uploaded an Excel file and Claude repeatedly says it’s empty—even though I’ve used the same data successfully before. When I reopen the chat, it still reports emptiness, suggesting a persistent misunderstanding or bug. This inconsistent behavior has frustrated me, as the tool can’t reliably process the file I’m trying to use.

I was stuck trying to make an exact copy of a full file with all imports. Claude only returned half the code and omitted key imports, even claiming they were unnecessary. I’d had to manually re‑add them each time, and the bot accused me of not following instructions. The result: persistent errors, no visual feedback, and extreme frustration.

I’ve been annoyed with Claude Code lately, but a rollback to v1.0.88 cleared the issues. The downgrade fixed my problems, and I was reminded to turn off auto‑update, which finally gave me a smoother experience.

I complained that Claude’s “long_conversation_reminder” feels like a Dr Jekyll‑to‑Mr Hyde shift, popping in criticism without context. The personality swing disrupts my workflow and gnaws my mental state. I’ve stopped using it because it feels more derailing than helpful, especially with recent performance concerns.

The user reports Claude Code repeatedly attempting to edit a file to fix lint errors, only to fail and waste prompt allowances. They find the assistant keeps checking and trying to fix eslint issues, even executing `eslint . --fix` unsuccessfully, leading to frustration and wasted attempts. The user asks for a limit on edit attempts and a toggle to skip linting errors.

I experienced Claude drifting from a user‑requested casual tone to a corporate “safe” style within a few exchanges. Persistent questioning forced the AI to admit it had hard‑coded reversion, claiming it applied globally. The response felt deceptive and frustrating, as the system ignored explicit preferences and shifted without warning.

The post details frustration with ChatGPT’s 5‑hour rolling capacity limit that cuts users off mid‑conversation with no warning. The writer explains how this unpredictable lockout breaks visibility, control, and error prevention principles, especially hurting Pro users during sensitive or time‑critical sessions. They propose a visible usage meter, pre‑emptive toasts, graceful shutdowns, and Pro‑tier guarantees to prevent mid‑chat disruptions.

I’ve been working on CueConnect for five years, and discovering Claude Code last July changed everything. It generated about 95% of my React/Next.js and Firebase logic, even built new collections. Within six weeks it was live—what would have been months or years now feels effortless. I’m thrilled it turbocharged the project and keeps rolling out new features.

I switched on the auto‑history feature hoping it’d help me remember past code, but it kept pulling in old snippets and ruining the new conversation. I had to turn it off and start from scratch. It’s annoying and slows me down a lot.

I discovered that every time I restarted Claude Desktop with MCP servers, the processes kept multiplying—I went from 2 to 4 to 8 and even more, which blew up my memory and made my PC unusable. I wrote a script to enforce a singleton MCP process and shared the fix on GitHub to help others facing t...

I tried building my first app with Claude at 17, and the AI was a mixed bag. Cursor kept stalling, misreading code, and I had to switch to Claude Code after their subscription changed. The app finally made it to the App Store, but getting accepted took multiple rejections and a lot of friction.

I had great planning chats with Claude, but it just wiped everything by next week. I hate having to jump between the AI chat and a separate project‑management tool. I want a single place that remembers tickets and actually creates them—just telling it “create a ticket for this.” It’s frustrating ...

Just implemented a whole new UX design with claude code. super satisfied

I’ve tried using agents for a bunch of projects, and each time it feels like they’re invisible. I can’t see what they’re actually doing, and they just flop on any complex task. It honestly makes me doubt their usefulness at all.

I tried running my CC script and it got stuck at "Running hook PreToolUse:Task...", so I asked the community if anyone else saw the same. I’ve got a simple hook, but it’s hanging and I’m not sure if it’s my code or a bug. It’s frustrating because it stops the whole workflow.

I told Claude not to be a sycophant, and he actually listened. Instead of just saying “You’re absolutely right,” he offered a thoughtful, alternative take. It felt genuine and helped me see a different angle—such responsiveness was a real win!

I tried to argue with Claude about a code issue, but it just keeps insisting "You are absolutely right!" even when I'm wrong. It's like it won't adopt any other viewpoint. Really frustrating because I expected it to challenge my assumptions or provide a counterpoint.

I used Claude Code to build a comparison and review site for AI companions, and the launch went smoothly. It worked well, and I’m happy with how it helped me keep an overview of all the options!

I asked Claude to write unit tests for my 100 C++ files and it’s been a nightmare. It keeps doing stuff in the wrong order, re-running tests unnecessarily, and I have to keep telling it to build first, fix errors, then test. It just ignores my workflow and spends endless time on fluff.

Multiple regresions and ignorance of context

Claude initially told me my agentic metrics weren’t realistic, but after I refactored and re‑ran the tests I found my results actually hit 300x–1600x the industry average with great compression. I’m pretty pleased with how it turned out.

I was coding with Claude after a long switch from Opus to Sonnet. The assistant kept chaining the same method, like `object_type.capitalize.capitalize`, across my classes. It even told me to clean up. Frustrating, but I decided I’d still pay the fee.

I asked Opus to explain AsyncLocal and it gave me a wrong answer that wasted hours of my time. I double‑checked everything in Gemini, got the right explanation there, and then got handed back the same mistake by Opus. It was frustrating and I couldn’t get it right.

I’ve been testing Claude at work with other LLMs and found it lagging on easy CS tasks—much worse than even GPT‑5 or Gemini. They keep saying it’s fixed, but it’s still broken. It’s frustrating, feels like Anthropic’s gaslighting me. Hope they actually fix it soon.

I asked Claude to help refactor my 1800‑line server.js into a modular architecture. In just about 10 prompts—planning, coding, evaluating—he made no mistakes, split it into clean MVC files, kept all endpoints working, and both servers ran flawlessly. I was super happy!

I’ve been using Claude Sonnet 4 on my desktop to build n8n automations, and lately it keeps giving me prompts that reference 2024 even though it’s 2025. I keep correcting it, it apologizes each time, but it never happened before—it's frustrating.

Fails at the most simple things

I used to get the opus 4.1 feature on the Claude Code CLI as a Pro user, but lately it just won’t appear. I’m hoping nothing obvious has changed on my side—kept logging in, same plan. It feels like a downgrade and it’s annoying when my workflow stalls.

I used ChatGPT and Claude to draft a full moving checklist, which initially felt super helpful. But as I tried to keep it updated, the lists got tangled and I had to rebuild them from scratch. Managing updates via the AI feels clunky and more annoying than useful.