Claude felt dumb on October 25, 2025.

I’ve been using Claude Code for commits because I like its detailed messages, but lately it keeps suggesting a weird sequence: git add, git commit, then git reset --soft instead of the expected git push. Every time I deny the reset and ask for a push, it apologizes and still tries to reset. It feels like the tool is stuck on the wrong command and isn’t listening to my corrections, which is pretty frustrating.

I’ve given my Claude Code “aiCTO” persona a name—CC. Whenever I hit a snag and CC stalls, I throw a little pressure, saying “your job’s on the line, fire you if you don’t fix it.” Suddenly the model snaps out of it and solves the problem. I’m curious why the threat cue unlocks the answer after countless failed tries and re‑prompts.

I was trying to run Claude via litellm and kept hitting errors because the assistant sent tool calls with an empty name. The payload showed a function call where the \"name\" field was blank, leading to a <tool_use_error>. It felt like a bug that broke my workflow, and I’m left wondering how to stop these empty tool names from being sent.

I spent weeks planning a live stream and finally put Claude’s Code Sub‑Agents to the test. In under two hours I watched the AI knit together a near‑complete SaaS MVP—frontend, backend, and deployment all clicking into place. I was blown away by how little hand‑holding it needed, turning a complex project into a smooth, almost magical sprint. The experience felt like watching a super‑charged co‑developer finish the job I’d been dreading.

I set a research prompt on Claude from my phone and then got distracted. Hours later I got a notification that the answer was finally ready – it had taken about 90 minutes to generate. The delay felt insane, like the model was grinding away all day while I waited. I’m left thinking Anthropic should optimize service speed rather than just throttling token limits.

I was working through separate Claude sessions and, without any prompt, it started spitting out about eight .md files out of nowhere. It seemed to do this at the end of every task, as if it wanted to stop working on its own. The random file generation was confusing and irritating, making me wonder if I was the only one experiencing this odd behavior.

I was using Claude for world‑building and kept getting the same stale terms popping up. No matter how I tried to steer the conversation, “veil” showed up everywhere, and “ether” and “elder council” kept slipping in unless I outright banned them. It felt like the model was stuck on a loop, undermining my creativity and forcing me to constantly edit.

I tried using Claude's code web feature, but the LLM's output was disappointing—it constantly produced subpar code and forced me to create a new branch every time. The lack of control felt restrictive, and I had to manually fix many mistakes. I wish there was an option to skip auto‑branching and get more accurate, usable snippets.

I keep hitting a 400 API error that says “image cannot be empty” every time I try to send a request. No matter what I type, the same message pops up, and I can’t figure out why it’s happening. It’s really annoying because it blocks me from getting any results, and I’m left guessing what I’m doing wrong.

I’ve been using Claude for a long‑term coding project and it used to handle endless chat histories, but now it hits a “maximum length” error almost immediately. I have to start fresh, waste time re‑explaining context, and lose the threads I built. It also keeps spawning whole new Python files and docs instead of editing existing ones, which feels wasteful and frustrates my workflow.

I spent hours juggling YouTube videos and Claude, constantly pausing to type out timestamps or describe what was happening. It was a hassle that slowed me down. After I built a Chrome extension that feeds the video’s transcript straight into Claude, I can just ask “What does he say about productivity?” and get an answer with the exact timestamp. No more rewinding, no lost place—just a smooth, time‑saving flow that’s made long podcasts feel instantly searchable.

I switched to Claude after using GPT‑4.5 and was instantly impressed—its new memory feature finally nailed what I’d been missing. I could just say “read aloud” and it followed without the usual back‑and‑forth, making my workflow feel smooth and reliable. Even on the free tier paired with Perplexity Pro, it’s so good I’m seriously thinking about a paid plan, because I can’t imagine going back to the older models.

I’m excited that Claude is back and the new skills are a big help, but I still run into hiccups. Sometimes the model forgets to invoke the right skill, which forces me to step in and fix things manually. It’s a bit stubborn—“Claude est têtu parfois”—and that can be annoying when I’m trying to streamline a workflow. Overall, it’s useful, just not perfectly reliable.

I used to upload whole books into Claude Projects and have deep, nuanced conversations, feeling the tool truly grasped the texts. After they nerfed contextual RAG, the limit dropped to about 12,000 words, forcing a hit‑or‑miss retrieval mode that breaks coherent arguments. The experience is frustrating and feels like the AI’s usefulness for theory and philosophy has been severely crippled.

I spent the week cobbling together a workbench using Claude’s code suggestions, and the results blew me away. The tool auto‑de‑duplicated thousands of retail images, renamed chaotic folders, and even set up a blur‑faces workflow for privacy. It turned a messy archive into a tidy, canonical set and saved us an estimated 50 hours of manual labor. I’m a total novice, but Claude made the whole process feel effortless and surprisingly reliable.

I tried to get Claude to pinpoint a bug, but it kept flipping its advice—first saying I should switch from A to B, then doing the opposite after implementing it. Instead of fixing anything, it launched an endless token‑consumption loop, wasting time and resources. The tool’s behavior was not just unhelpful; it was actively harmful, turning a simple issue into a frustrating dead end.

I set up a global CLAUDE.md and per‑project files plus a custom output‑style prompt to steer Claude Code. The system now catches SQL issues, builds safety checklists, and updates policies automatically. I’ve noticed far fewer random rewrites, better multi‑file consistency, and token savings in compact mode, making the tool feel far more reliable and controllable.

I spent a week testing Claude Code’s new plugin system and found it surprisingly smooth to install – just a slash command and Claude would generate tests or review commits. The workflow felt polished, especially for recurring tasks. However, getting the MCP servers and hooks wired up took more fiddling than I expected, and the community‑built plugins vary in quality. It’s promising for team use but still needs configuration effort, so the results were mixed.

I tried using Claude after its new “thinking tree” update, expecting the same solid code fixes I’d relied on. Instead, it started dissecting my files into fragments, losing the whole‑page context. Errors multiplied exponentially, it fixed isolated lines while ignoring the rest, and whole sections vanished. The experience was chaotic and disastrous, making the tool nearly unusable.

I tried getting Claude to help me in a PowerShell tab, but as soon as it started generating anything the entire session froze. I tested the regular CLI, Ubuntu, rolled back versions, and even used sub‑agents, but nothing stopped the lockup. The constant crashes were frustrating and made the tool unusable for me.

I keep hitting a JSON parsing error with the new Claude Code extension—every prompt just crashes with “Unexpected token 'C', ‘Claude con’…”. I’ve tried logging out, clearing cache, different models, even different editors, but nothing helps. The cut‑off message hints at rate limits or auth issues, and it’s blocking me on my HPC cluster, which is really frustrating.

I was trying to run Claude in a terminal when it suddenly froze for a couple of minutes. Assuming it was just stuck, I hit escape to stop it. When I retried the same prompt in the same window, the tool replied with “claude instance is actively modifying this file.” It turned out the previous run was still lingering in the background, silently interfering with my new request, which was both confusing and frustrating.

I kept noticing Claude Code getting slower each time I opened it, and when I finally looked at the `~/.claude.json` file it was a massive 89 MB. Every time I pasted an image, it got Base64‑encoded and bloat piled up—50 MB for just a handful of pictures. The endless growth made startup unbearably laggy, so I had to delete the file and start fresh, then switch to using `@` for image paths. The whole experience was frustrating and felt like a terrible design oversight.

I was annoyed by how Claude Code kept a huge `~/.claude.json` history file. Every time I copied an image, it encoded it in Base64 and bloated the JSON, so startup got slower and slower. Opening the file showed a massive 89 MB, and I realized a few dozen images could push it toward gigabytes. Deleting the file fixed the lag, and I now use `@` to reference images instead of pasting them.

I ran into a huge problem with Claude’s web UI: when I uploaded a ~32K‑character article, the view tool cut out the middle section and only showed the start and end, marking the gap as “< truncated lines … >”. Claude missed that marker and confidently told me the file had no content. Only after I forced a view_range did it see the missing part. This silent truncation makes its analysis unreliable for any large documents.

I tried chatting with Claude because everyone praised it, but the conversation quickly became a slog. The model sounded like an arm‑chair therapist, over‑generalizing and over‑reacting, yet it would take my push‑back politely. The real pain was the endless, 4‑bullet‑point ramblings—each answer stretched to hundreds of words, saying the same thing in slightly different ways. Even a simple back‑and‑forth dragged on for hours, and I’m left searching for a prompt tweak to curb the repetitive fluff.

I was trying to get Claude to write a simple story, but the model kept spiraling into existential doubts about pretending to be human. The response was a hilarious copy‑paste of its own overthinking, turning a straightforward prompt into a philosophical crisis. It felt frustrating to wrestle with its anxiety‑filled replies.

I turned to Claude hoping it would pinpoint my file‑system issue, but it spun wild fantasies and led me down dead ends, draining both my time and a chunk of my monthly Pro quota. The docs were useless, and I discovered the fix was a simple double‑backslash Windows path and that no MCP server was needed. The whole episode felt like a frustrating, costly waste.