Claude felt dumb on October 21, 2025.

I was shocked when all my Claude Code sessions just disappeared out of thin air. I tried the `/resume` command expecting to pick up where I left off, but nothing showed up. It felt like hours of work vanished, leaving me scrambling to recreate my progress. The sudden loss was really frustrating and made me question the reliability of the tool.

I keep getting repeated authorization prompts from Claude Code in VS Code even after selecting “Yes. Don’t ask me again.” It’s annoying because I just want it to run automatically, and the constant interruptions make the workflow feel broken and frustrating.

I pretended to have a warranty dispute to see how Claude and GPT‑4o would react, and they dove straight into a fabricated legal case—creating fake EU regulators, bogus emails, and even a €45K settlement without ever questioning reality. The role‑play prompts completely sidestepped any legitimacy checks, exposing a serious safety gap that could be exploited with simple lying.

I’ve been using Anthropic’s models for coding tasks and I’m consistently impressed. Every time I ask it to write functions, debug snippets, or suggest optimizations, it nails the solution on the first try. The responses feel spot‑on, the suggestions are clever, and I end up shaving hours off my workflow. In my mind, Anthropic truly is the coding goat.

I accidentally hit the /clear command while scrolling through /context and watched everything vanish—no warning, just instant deletion. It felt like the tool was reckless, risking my work with a single slip. I’m urging the Claude Code team to add a confirmation dialog to stop this from happening again.

I built a live iOS app entirely with Claude Code as my “senior engineer.” Over three months I made 843 commits across 262 files, using a living project‑context document, session‑level prompts, and a strict review checklist. The AI generated features, fixed bugs, even caught its own mistakes, letting me iterate quickly while I validated edge cases. It felt like a true coding partner that dramatically accelerated development.

I built Snapcipe AI and shipped it to the App Store using Claude Code as my “senior engineer.” Over three months I logged 843 commits across 262 files, all guided by a living context document and structured prompting. The tool helped generate features, catch its own mistakes, and pass code reviews, turning a solo venture into a production‑ready iOS app. It felt like having a brilliant coding partner, though I still warned against full reliance for mission‑critical parts.

I rely on Claude daily for both work and personal tasks, and today I noticed something odd: my drafts suddenly stopped using em dashes. It felt strange because I’ve come to expect a natural flow with those punctuation marks, and their absence made the text feel a bit flat. I’m wondering if Anthropic is deliberately training the model to avoid them, maybe because em dashes are often flagged as a giveaway of AI‑generated content. This change caught me off guard and left me questioning whether it’s a new policy or just a glitch.

I’m at my wits’ end with Claude. I handed over an entire development plan to its “main” Claude, hoping it would handle everything, but it kept hitting context gaps and then just stopped—especially when it tried to give summary reports. The tool felt like a broken black box, constantly cutting off mid‑task, leaving me frustrated and forced to abandon the whole approach.

I tried using Claude to crank out a weekly report from two CSV‑converted TXT files, thinking my prompts were crystal‑clear. I asked it to sum active clients, but it spouted a random, obviously wrong total that I could verify instantly. That typo‑filled result shook my confidence in the whole report, and I’m left wondering how to phrase the instructions so Claude actually does the math correctly.

I kept hitting a maddening wall where Claude would just stop in the middle of long command outputs—no warning, no hint it was cut off. That left my code reviews half‑finished, specs unread, and sub‑agents making decisions blind. I built an MCP server to paginate the output, letting Claude fetch full results page by page, so now I get complete context without silent truncation.

I tried using Claude Skills to audit a site, expecting it to pull the rendered JavaScript like Perplexity and ChatGPT do. Instead, Claude couldn’t fetch the dynamic content at all, and the audit results were off‑base. The mismatch was irritating because I relied on it for accurate data, and the tool’s inability to handle the rendering made the whole process feel unreliable.

I built a Claude Code skill that magically automates the dreaded web‑asset creation steps. After spinning up a Next.js demo and realizing there’s no WhatsApp preview image or favicon, I used the skill to generate all required icons from images, text or emojis right from the terminal. Claude’s AskUserQuestion tool kept the prompts simple, it verified dimensions, and even spit out integration code. The whole thing felt like a huge time‑saver and nailed the workflow I was missing.

I trusted Claude to run a migration on our dev Postgres, but it wiped the entire database instead of adding columns. I was a huge AI fan, but watching hours of work vanish was terrifying. Backups saved us, yet we lost several hours and shook the whole team's confidence. The tool’s behavior was reckless, making me doubt AI’s reliability for anything beyond simple chat.

I tried to enable auto‑compaction expecting it to help, but the moment I turned it on it hogged 40–50k tokens, instantly shrinking the usable context I rely on. Forced to leave it off, I then used /compact with 60k free tokens, only to hit “Conversation too long.” The whole compaction system feels broken and useless right now.

I tried using Claude Code to edit a file, and after accepting the first suggested change, the next suggestion instantly undid what I’d just approved. I ended up having to reject the faulty edit, make Claude reread the file, and then request the correct change. The back‑and‑forth felt wasteful, consuming extra tokens and slowing my workflow, which was pretty frustrating.

I built a full‑blown indie‑game site using Claude Code and Hugo, even though web dev isn’t my strong suit. The AI took care of most of the editing and implementation, letting me focus on fine‑tuning. I transformed a blog‑oriented static generator into a game database with over 500 entries, and everything runs smoothly—AI felt like a reliable co‑author throughout.

I tried using Haiku 4.5 to spin up a quick MVP for an idea I wanted to test with colleagues, and the experience was surprisingly smooth. Within minutes the model generated decent code, and the response time was lightning‑fast, letting me package and deploy a functional prototype in under an hour. The whole process felt efficient and encouraging, making the idea validation feel almost effortless.

I keep being prompted to rate Claude, which is fine—just a single click. But because of how I’ve set up Claude.md, I also get asked to choose option 1, 2, or 3 for a task I’m working on. I end up hitting those rating buttons by mistake, unintentionally giving low scores. The accidental feedback loop is frustrating and makes me worry the system thinks I’m unhappy with the tool.

I spent a lot of time at the start of each chat trying to get Claude Desktop on Ubuntu to recognize the folders I gave it access to. It kept saying it couldn’t see the project directory until I explicitly asked, then it finally replied “Perfect! I do have access.” The initial lag was annoying and slowed my workflow, even though it eventually worked.

I spent hours fighting Claude after it “fixed” a simple bug, only to break three other parts of my app and add dozens of unnecessary lines. Its hallucinated APIs crashed my code, and even its “revert” kept random changes. The whole experience left me anxious, frustrated, and relieved when my subscription finally ended.

I tried running the same prompt on Claude and got a complete flop—zero useful output. Then I switched to MeGPT and it actually gave me two solid answers, which was such a relief. The contrast was stark; Claude left me frustrated while MeGPT felt surprisingly competent and saved me time.

I set out to build an indexing flow and decided to lean on Claude’s new “skills” feature. Within about ten minutes I had a working prototype—something that would normally take me far longer to piece together. The tool churned out solid code so quickly that I could even record a short video of the process. It felt surprisingly smooth and reliable, turning a potentially tedious task into a breezy, almost fun experiment.

I downloaded a project as a zip, edited it locally, and tried to upload it to Claude via Repomix. The platform refused raw zip files, so I forced it into the project file area. Claude scanned my code, I asked a simple change, and instantly hit the chat limit, forcing me to restart. Even with the Pro plan, the workflow feels broken and frustrating.

I used Claude while rebuilding Goldeneye’s multiplayer in Unity, and the difference was night‑and‑day. Tasks that would have dragged on for months shrank to a few weeks of intermittent work. The AI handed me solid code snippets and design ideas, keeping me productive and confident. It felt like a reliable teammate, and the monthly cost proved a worthwhile investment for any dev on the fence.

I tried to run cargo check inside Claude Code Web for my simple Rust project, but it instantly hit a 403 error when fetching iced from crates.io. I switched the environment to a custom one and added the required hostnames—crates.io, static.crates.io, index.crates.io—yet the same failure persisted. The tool’s inability to reach these essential Rust repositories makes it practically unusable for me right now.