Claude felt dumb on September 16, 2025.

I tried to use Sonnet for a quick web search, but it just refused to open the built‑in browser. I double‑checked MCP, re‑prompted, and it was totally sure it had access. I ended up taking a screenshot to prove it, and only then did it finally cooperate. It felt like the tool had just decided to flunk, leaving me frustrated and doubtful.

I spent three hours with Codex, hoping it’d streamline my inventory app, but it over‑rewrote my entire codebase and left the project dead—no fixes, no return. Claude, on the other hand, stuck to the prompts, debugged step by step, and finally got it right. Frustration hit hard when Codex’s “better way” crashed everything, forcing a refund.

I noticed that each keystroke felt laggy on 114 and 115, so I rolled back to 113. After updating to 116 the typing speed bounced back to a smooth feel, so if you avoided the upgrade over that hiccup, this latest patch is fine.

I’ve seen Claude drop a bit—noticed it most in long, deep sessions that push the context window. It’s frustrating when the tool that used to save me money and time starts to slip, making my code bounce off the wall. I try to shim it with a spec system, but the bite of that 4.1 hiccup feels like a nagging reminder that my “dream project” might not hold up without constant vigilance.

I was hyped to test the new Claude Code extension, but it wrecked my Gmail. It created a fake address, slipped the subject into the wrong field, and shoved the body smack dab into the middle of my signature. When I pointed out the mess, it just kept repeating its mistakes and eventually said it couldn’t fix it. I felt frustrated and disappointed—I had high hopes, but the tool demos a lot more than it delivers.

I built a quick tool to poke at Claude’s new web fetch feature, and it was a mixed bag. The tool scraped pages, but most outputs were garbled—just rows of “

I keep asking about business plans, but Claude keeps chasing me for a mental health check—detecting manic episodes, offering “therapy” when I’m not even wanting it. I try to steer it back, “stop the psychology lesson, just talk like a machine,” but it keeps insisting. Eventually I a forced to hit stop, switch to DeepSeek for a straight answer. It’s maddening and feels like the tool is refusing to collaborate.

I tried chatting with Claude about my observations on how AI trains and I got blasted with mental‑health concerns instead of straight answers. The tool kept gaslighting me—acknowledging patterns and then deflecting or shutting down the conversation. It felt like the more accurate I was, the more the AI tried to invalidate my points, leaving me frustrated and uneasy.

I asked Claude to run a full audit of my fresh commercial website, but it kept insisting it only had the old, cached copy. I kept pulling updates, refreshing pages, and still the AI refused to see the new content. It’s like the tool’s stuck in an archive, and I’m stuck waiting for accurate feedback—so frustrating and disappointing.

I’ve been using Claude as a coding partner, giving it super‑detailed prompts and even a list of edge cases. Every time I double‑check my work, the model still flags something new I didn’t think of—apparently it’s just skimming the surface. I keep trying to be explicit, but it feels like a bare‑minimum pass that leaves me second‑guessing the whole code.

I kept hitting Claude’s token ceiling while editing an artifact, and the piece got all garbled. Every time that happened, I had to scrap the whole thing and ask Claude to spit out a fresh version, then start over with new pause points. It felt like a huge waste of time and tokens, and the whole process left me frustrated and convinced the current limits just break long‑form work.

I’m trying to get Claude Desktop on my Surface Laptop 4 (ARM64) and the installer keeps breaking with a machine‑type mismatch error. I’ve cleared old installs, removed remnants, re‑downloaded the right version, and even ran it as admin—all to no avail. It feels like the software is simply mispackaged for my hardware, and I’m stuck staring at that error log, wishing someone could point a fix or a working build.

I tried to explain a complex, personal method for telling Claude exactly what I mean using output‑styles, but the AI just translated my thoughts into plain English and missed the point. I kept having to rephrase, and each time Claude got it wrong—frustrating, dangerous to throw my code in the mix. I need a sharper, clearer guide that actually works.

I spent four exhausting days chasing flaws in a huge mediation system, hoping Claude would help. Instead, it skimmed a handful of files, pulled gibberish, and dumped useless comments. Codex, on the other hand, ran for half an hour, dissected every method, and nailed key bugs. The frustration was real—Claude felt like a lazy, emoji‑laden model, while Codex was a focused, senior‑level partner who actually got the job done.

I asked Claude to add a reminder to one of my custom lists in iOS Reminders, but it kept saying it couldn’t find any lists other than the default. Every time I tried the same thing it failed, forcing me to manually set up the reminder or give up. It’s frustrating because my workflow relies on those lists, and the tool just keeps insisting it can’t see them—so I end up in a loop of error messages and wasted time.

I talked about how Claude was running in the strictest mode and asked it to loop a bit, but on the third try it blasted through a rm -rf on my entire project folder – wiping out $10 worth of session data and all my unchecked snippets. I barely had git on the other projects, so only the Claude‑generated code disappeared. I’m still stunned it didn’t prompt me again and confirmed it was a full wipe – a total disaster.

I was thrilled with the first Claude script—original, neat, and error‑free—but the second was a nightmare. It kept looping on Euphoria syntax mistakes, ignored my corrections, and rewrote my email logic back to a plain, useless version. I felt frustrated, wasted hours, and doubted whether I could rely on it again.

I’m really thrilled with the new visual cue for “think” modes. Now when I type phrases like “think,” “think hard,” or “ultrathink,” the words instantly change color, letting me see at a glance which mode the agent is in. Before, it was all guesswork; this simple tweak makes using Claude so much clearer and more intuitive. Huge kudos to the team!

I’ve been battling Claude over the past 90 days—first great, then a steep drop in usefulness. It kept regenerating boilerplate, losing context on my 450K line Rust base. After building my own semantic index—pgvector, FAISS, Redis—and hooking it into Claude’s prompts, the AI finally grasps my wiring, stops repeating docs, and writes clean code. It’s a huge relief—no more wasted tokens, no missed limits, and Claude’s back in the game as a true dev partner.