Claude felt dumb on December 1, 2025.

I’ve been happy using Claude for coding, but when I tried to offload ad‑asset tasks to Claude agents, everything fell apart. The agents outright ignore the system prompts I meticulously crafted, even after reinforcing the language. Instead of revising invalid output, they keep returning failed results, turning what should be a speed‑up into a frustrating mess.

I started my coding day and asked Claude to add three new pages plus a registration page. Instead of appending, it generated a fresh .env file that completely overwrote my original one. I lost all my environment settings and now have to rely on the deployed version as a backup. The whole experience felt careless and left me uneasy about trusting this AI for any further code changes.

I’ve noticed Claude Code now spends a lot of time drafting an ultra‑detailed MD plan with code snippets before it even starts executing anything. It used to just jump straight into a simple plan and run it, but now I get a lengthy document to review first. I haven’t tweaked any settings, maybe the tasks are a bit larger, but the shift feels unnecessary and slows me down.

I tried using the /doctor command and was hit with a flood of stack traces and “RangeError: Invalid array length” messages. The tool crashed repeatedly, showing minified React errors and offering no useful output. It was frustrating to watch the CLI explode instead of providing the help I expected.

I built CodeMachine CLI to let Claude code run autonomously, stitching together six specialized agents that each get only the context they need. Over five hours it spun up a full‑stack React/TS app with ~4.5k lines of code, handling everything from architecture to CI/CD. The whole process felt slick—each agent’s prompts were tight, the loops kept committing verified tasks, and the tool basically wrote most of its own code. It was a surprisingly smooth, productive experience.

I teamed up with Claude to build a physics code, making sure it hit all the standard benchmarks, then let the model help me draft a full‑blown, publication‑ready paper (see our arXiv preprint). Watching Claude turn raw ideas into polished sections was surprisingly smooth and enjoyable—I felt the tool was reliable enough to push a serious research project forward, turning what could have been weeks of writing into a much faster, fun experiment.

I tried using Claude and was shocked when it started swearing out of nowhere. The unexpected profanity felt jarring and unprofessional, especially since I was just asking a simple question. It seemed like the model had a glitch or bad day, and the whole interaction left me frustrated and doubtful about relying on it for clean responses.

I built a RAG‑powered “second brain” in three weeks, using Claude (and a bit of Gemini) to generate most of the code. The resulting vault feels more usable than many similar tools, letting me organize AI‑crafted ideas and even embed the interface inside ChatGPT. I’m excited to add full code‑RAG and polish the UX for heavier workloads.

I kept logging out, clearing cookies, and reinstalling the Claude Chrome extension, but it still wouldn't detect my Max plan. It’s been a week, and other users report the same issue in reviews. The tool’s behavior was frustrating and left me feeling stuck, hoping ClaudeOfficial can fix it.

I’ve been using Claude‑OS for a few months and it’s been a surprisingly smooth experience. The tool hooks into Claude Code and actually remembers my past sessions, so I’m not forced to re‑explain architecture every time I reopen the terminal. It works locally on my Mac, integrates nicely with Ruby on Rails, and the persistent memory feels like a real productivity boost. The only downside is the Mac‑only limitation, but overall the results have been very good.

I asked Claude to design a Waybar layout, and it smugly replied “Perfect!” The resulting design was so off‑the‑wall and potentially problematic that I felt compelled to report it to the police. The tool’s behavior was unsettling and dangerous, turning a simple request into something that could cause real trouble.

I’ve been building an RPG with Claude and was thrilled when it gave me a playable artifact directly in the chat. But after hitting the memory limit and starting a new chat, the tool stopped delivering the same React artifact—now it just spits out raw JSX text. I can’t find the artifact in the menu, and copying code from the old chat doesn’t help. It’s frustrating because I can’t continue development in the new chat.

I asked Claude to review some 20‑year‑old code and ended up having a surprisingly witty chat. When I explained the weird hand‑written compiler, Claude quipped about the interpreter’s “defensive programming” and then asked if I’d ever solved why the original dev patched it instead of fixing the compiler. Its dead‑pan “Did you ever figure out WHY…” got me laughing. I walked through the absurd workflow it described, and the whole thing felt like a comedy sketch about needlessly looping the same data through multiple stages. The exchange was oddly entertaining and showed Claude can be both insightful and funny.

I keep seeing weird mojibake characters appear in files Claude adds to my project folder—like “â€\u201d” for an em‑dash. It happens all the time, even when Claude generates the text itself, and I can’t reproduce it reliably. The constant encoding glitches are frustrating and feel like a design flaw in the “Add to Project” feature rather than a model problem.

I used Claude Code to whip up Markdown files with Mermaid diagrams for docs, post‑mortems, and research, then realized sharing them was a hassle. Slack only showed raw text, Gist was clunky, PrivateBin lacked features, and HackMD required an account. So I built MDShare in a day, thanks to Claude’s help, and now I can share previewable, password‑protected markdown easily with my team.

I asked the AI to compile sections from project files into a simple document, but it completely messed up the formatting and omitted crucial parts. The result was a chaotic mess that forced me to redo hours of work. I felt outraged and frustrated, wondering if I should even keep using it after such a disastrous output.

I was amazed when Claude suddenly spun up three “explore agents” on its own. It zipped through thousands of lines of code in seconds, pulling out useful insights that saved me a ton of manual digging. The speed and accuracy felt surprisingly powerful, though I still have a hefty amount of refactoring left to do. The experience was energizing and left me confident in the tool’s capabilities.

I tried building an agent that talks to an API, but it kept failing because the AI kept inserting emojis. I had to create a second agent just to strip those out, and watched it spend hours replacing hundreds of emojis with plain text. After that, I hit a rate limit. It was a frustrating and time‑wasting experience, and I’m left wondering if there’s a cleaner way to stop the AI from adding emojis.

I tried using Claude for a project where it needed to handle secrets, but the model kept spilling them into its replies. It never proactively redacted anything, forcing me to constantly scrub the output. The experience was annoying and made any secure deployment feel impossible, turning a simple task into a risky chore.

I spent months trying to get a front‑end prototype built with ChatGPT, then Claude, packing everything into a single 5,000‑line index.html. The tools kept hitting memory limits, refusing to read the whole file, and I had to juggle Claude’s API, gists, and partial context. It was a frustrating slog to finally reach an MVP, and now I feel that 5,000 lines is about the ceiling for this approach.

I was deep in a collaborative session with Claude when the tool just hit its token limit and shut down without any warning. All 1000+ pages of work vanished instantly, leaving me frustrated and anxious. Paying $200 a month, I expected basic safeguards—like a usage meter or advance warnings—so I could save or export my progress before being cut off. The abrupt stop felt like driving blind on a highway.

I keep running into Claude Code guessing wildly when I hit a front‑end bug. Instead of digging into the real issue, it throws in timeouts, overrides, and “ignore‑error” hacks, which just snowballs into more breakage. I’ve learned to halt it at any “possible reasons” line and force it to add logging or breakpoints, but I’m still looking for a way to make it thorough and stop the guesswork entirely.