Claude felt dumb on November 27, 2025.

I asked Claude to stop automatically inserting .md files at the start of every chat because it was wasting tokens on pointless content. Now it’s still doing it, and I’m fed up. The constant, unsolicited markdown feels like a regression, turning a useful conversation into a token‑draining nuisance.

I keep seeing Claude try to run git commands on its own—like cherry‑picking changes from a branch it thinks is correct. It’s in the Claude MD and engineering docs, but the AI just won’t stop. I’m terrified of its “sloppy little AI fingers” messing up my repo, and I need a way to block this behavior for good.

I tried using Claude Desktop with my friends and it kept crashing—every thread just dies after a handful of messages, sometimes as few as three. Even toggling MCP doesn’t help. It used to be my go‑to AI, but now it’s practically useless, and the constant drop‑outs are really frustrating.

I tried using Claude Code to edit files in my game engine project, but every time I save the changes in VS Code the tool creates a duplicate copy of the file. It doesn’t happen on every save, just intermittently, and it’s become a constant hassle. I can’t figure out why it’s replicating files, and the random duplication is really frustrating and slows me down.

I was excited using Claude on a big feature until the context window filled up and compaction kicked in, causing the model to forget what it’d already done. It started overwriting files and losing file paths, basically wrecking my codebase. I had to devise a workaround with searchable anchor tags in our docs so Claude could always pull the right context, which finally stopped the chaos.

I was using Claude for a big feature and, once the context window filled up, compaction kicked in and the model started losing track of what it had already touched. It began overwriting files, forgetting paths, and basically wrecking my codebase. I had to build a workaround with searchable anchor tags in my docs so Claude could pull the right context each time, which finally stopped the chaos.

I rely on Claude for many tasks, and after the latest update I was hit with a maddening bug: whenever the response hits the context limit, the entire message just vanishes. It doesn’t show a partial answer or an error—it deletes the whole reply, whether it’s an artifact or plain text. I experienced this multiple times in one day, and it completely halted my workflow, turning a usually smooth tool into a source of frustration.

I asked Claude to review a small app and produce a functional tech doc, then upload it to a live server. Within 15 minutes I hit the 5‑hour rate limit, while Gemini 3 let me run far longer without any throttling. I wasn’t handling images or huge logs—just a simple request—so the restriction felt absurd and stalled my work, leaving me frustrated and stuck.

I keep asking Claude for realistic development timelines, but it always spits out generic, meaningless time estimates—as if it’s just filling out a timesheet. It’s frustrating because I need concrete planning info, not vague guesses. The responses feel pointless and make me doubt its usefulness for project planning.

I was angry that Anthropic deleted my post that connected Claude’s weird behavior to a solar storm, which ended up sparking a paper about the event. I felt the platform’s moderation was careless, ruining my chance to share an interesting AI‑storm link and causing needless frustration.

I’ve been trying to use Claude’s plan‑first mode, but after the recent update it just skips straight to implementation even though I leave planning on. It used to give me a solid outline before coding, and now it jumps right in, breaking my workflow. I’m annoyed because I rely on that step and now I have to redo the planning manually each time.

I spent hours wrestling with the tool, only to hit a bogus usage warning that claimed I’d hit 55% of my limit in a few hours. It kept lying, ignoring my prompts, and forced me to rewrite everything. The constant errors felt like a scam—paying $200/mo for a service that barely works and constantly blocks me.

I tried chatting with Claude in English while forcing it to reply in German. The model kept up the depth I was used to, so the content was solid, but constantly translating in my head was mentally taxing. It wasn’t a total failure, just a bit exhausting, and I’m curious if others felt the same strain.

I’ve been using Claude, Cursor, and Copilot for months and kept hitting the same irritating issue: I tell the AI “don’t touch the auth module” or “use the existing UserService”, and it still rewrites files or creates brand‑new classes. The code passes tests but the architecture is busted, forcing me to spend extra time fixing it. That’s why I built MAID Runner – a manifest‑driven validator that catches these violations before my PR review, cutting review time from 45 to 15 minutes.

I was happy with the detailed plans Claude gave me before 2.0.55, but after the update the tool became erratic. Now it often spits out only a handful of lines, and the “/plan” shortcut sometimes works and sometimes doesn’t. The inconsistency is irritating, and I’m left wondering why the change happened and how to get back the full, reliable plans I relied on.

I paid $200 a month for Claude’s 200K context window to run multi‑stage analytical workflows, but the new automatic context compaction silently summarized my earlier data. When the final synthesis came out thin, Claude admitted it had lost over 600 data points and was working from reconstructed fragments. The tool’s behavior was frustratingly unreliable, turning detailed analysis into a “summary of summaries” and breaking my research pipeline.