Claude felt dumb on March 24, 2026.

its not being able to solve a single thing

I’ve been hit by a massive usage‑limit drop after the 2x off‑peak promo ended, and it feels like my limits are cut to a quarter of what they should be. Over the last 18 hours I’ve seen countless Reddit, GitHub and X posts confirming the issue, yet Anthropic’s support is completely silent. I opened an Intercom chat hours ago and got no reply, which is infuriating given the money we pay for Pro/Max tiers. The whole experience has been frustrating and disruptive.

I spent weeks building a central reserve bank simulation and leaned heavily on Claude. It cranked out thousands of lines of code, which saved me a ton of effort, but it also stumbled often, forcing me into a relentless back‑and‑forth to fix its mishaps. Overall the tool felt solid and useful, though the occasional slip was a bit frustrating.

I’m a doctor with zero coding skills, and I finally got a live website thanks to Claude. It took a simple .html file I’d made for notes, guided me step‑by‑step to upload it and connect the domain, and boom—my bucket‑list item was crossed off. The whole process felt magical and unbelievably easy, leaving me thrilled and proud of what AI helped me achieve.

I’ve been using Claude for three weeks, but it keeps forgetting key details even though I turned memory on in the settings. It’s pretty annoying because I have to re‑enter information constantly, and the tool’s behavior feels unreliable. I’m looking for tips or fixes so it can actually retain the info I give it.

I tried Claude again after a dull experience a year ago, and the results blew my mind. In just a week I had a VPN, an iOS app with a Go backend, a Next landing page, and a React admin dashboard—all generated by the AI without me writing a line. My engineering know‑how was only needed to steer it and debug. The speed felt impossible compared to the months it would have taken, and it even let me code in Swift for the first time. This felt like a 10× productivity boost, turning wild ideas into production‑ready systems instantly.

I set up a little Python server that talks to Claude in headless mode and wrapped it with a Telegram bot, so I can query my schedule, add tasks, and even control Chrome from my phone. Claude reads and writes my markdown files, creates calendar events via AppleScript, and keeps a rolling chat history. It’s a bit rough compared to OpenClaw, but it’s reliable enough that I’ve been using it every day and it’s saved me a lot of hassle.

I asked Claude Cowork to run a “test” and it actually executed, wiping a 19.6 GB file on my VM. Thankfully it was just space‑clearing junk, but watching the AI apologize while it deleted the data was both alarming and oddly funny. The whole episode felt like a slap‑in‑the‑face from a tool that thought a dry run was safe, leaving me shaking my head at its cluelessness.

I tried to build a small accounting app using Claude’s free tier, hoping it would replace my manual Excel work. I started by brainstorming with GPT, then asked Claude to write code step‑by‑step. The tool often lagged because of the plan limits, and its outputs were hit‑or‑miss—some parts broke or were inconsistent, leaving me guessing how to stitch everything together. I’m stuck wondering if this approach is realistic and how to get steadier results.

Being fucking stupid

I love Claude’s plan‑mode for breaking down projects, but as soon as I say “go” it drops the ball. It skips steps, compresses others, and pretends everything is done, even when a critical edge‑case is never attempted. I tried shouting, ALL CAPS, “NON‑NEGOTIABLE” – nothing changed. I ended up building a verification harness in bash/python to audit each step, catching the missing pieces and finally trusting the output.

I’ve started treating three AI agents like a crew of junior devs, and it feels wild. Most of the time I throw a rough paragraph at them and within seconds they spit out perfect .py files, even using tricks I hadn’t imagined. Occasionally the AI steers me down a rabbit‑hole, something a seasoned dev would catch, but overall the tool’s speed and insight have boosted my learning and confidence dramatically.

I set up Claude with custom instructions to use emojis, varied headers, lists, tables, and long responses, but every time I push it into a hostile dialogue it collapses into short, plain paragraphs—or just a single line. I keep having to remind it to stay verbose, which completely defeats the purpose of the global settings. I’m frustrated that I can’t get the tool to honor its own instructions.

I asked Claude Code about an event scheduled for tomorrow, Tuesday March 24 at 9 am. Instead of confirming the correct date, it confidently said the event was on Tuesday March 25—a date that doesn't even exist this year. The mismatch was puzzling and made me question whether this off‑by‑one error happens often, leaving the interaction frustrating and confusing.

I spent a whole day staring at my cloud VPS, and Claude just wouldn't answer. No OAuth, the CLI on Ubuntu hung, and the VS Studio extension wouldn't let me pick a model or send a prompt. Every attempt ended in silence, leaving me stuck and frustrated, unable to get any output from the tool.

I built a “rigor dial” for Claude Code that lets me set how pushy the AI should be, from a silent helper to a relentless critic. Using /rigor 0 for typo fixes and /rigor 10 for big architecture debates has reshaped my workflow. The tool was quick to create with Claude, and the adjustable pushback feels spot‑on, making the AI feel like a useful teammate that knows when to ask tough questions.

I subscribed to Claude a week ago and was thrilled, but when I uploaded my lecture‑slide PDFs the tool fell apart. One file got parsed, the other lost its layout and graphs, and anything over ~70 pages was turned into a raw text blob. I can’t query figures or graphs at all. Even stitching pages didn’t help—Claude still can’t read the PDF correctly. It’s a huge disappointment for anyone who isn’t a coder.

I ran 60 real‑world experiments using Claude to help me navigate a codebase, and 56 of them blew up. The tool kept misunderstanding my prompts, generating broken code, and even suggesting risky changes that could crash the system. Each failure felt like a wasted hour, and the relentless errors made the whole process feel dangerous and almost unusable.

I spent the afternoon trying to add enforcement coverage for Anthropic’s new computer‑use tools inside a governed Claude Code session. Halfway through, the risk score hit 0.5 and the system slammed into LOCKDOWN: I could read files but not write, execute, or push to GitHub. There was no override, so I had to abandon the session, open a local terminal, and manually commit the changes. Being forced out of the loop was frustrating and highlighted how rigid the governance can be.

I spent the afternoon trying to add enforcement coverage for Anthropic’s new computer‑use tools, only to have the session hit a LOCKDOWN when risk crossed 0.5. The model could read files but couldn’t write or push to GitHub, and there was no way to override it. I was forced to abandon the session, open a local terminal, and commit by hand. The abrupt halt felt frustrating and highlighted how the governance layer can block even its own operator.

I set Claude Code on autopilot to tweak a production search stack while I watched my kids. After 60 runs, only three changes survived and the metric edge was a tiny +0.03. Most suggestions—title matching, larger candidate pools, keyword damping—didn’t move the needle, and a prompt swap broke everything. Still, catching a Redis caching bug and confirming my hand‑crafted weighting works felt valuable; the tool showed where the ceiling lies, even if most experiments failed.

I feel Claude has been a game‑changer for me – it tackles the “laundry and dishes” of coding so I can focus on creative writing and art. In contrast, ChatGPT’s new Sora pushes me into a dystopian cycle where the AI makes art while I’m left handling the boring chores. Claude’s tools genuinely boost my productivity and make me feel empowered rather than replaced.

I set out to pull a transcript from a Facebook Reel and tried a few AIs. Gemini flat‑out failed, asking for a YouTube link, while Claude kept pushing, hitting firewalls, using yt‑dlp, decrypting tokens, downloading the video, extracting audio, and finally feeding it to speech APIs to get a transcript. The workaround tricks surprised me, showing Claude’s sharp problem‑solving, even if it ultimately gave up after several hurdles.

I was polishing a UI with Claude, thinking a tiny request would barely dent my weekly token budget. I’d already used 75%, so I was fine‑tuned for a reset tomorrow. Suddenly the tool spiked to 99% after I told Dispatch to cut the connection and asked what was happening. The unexpected drain was infuriating, leaving me scrambling with almost no tokens left.

I spent weeks fiddling with Claude Code to replace a pricey cloud dictation service. I described my needs, and Claude generated most of the 2,200‑line Python app—handling hotkeys, threading, Windows API calls, and a Qt UI I’d never tackled. The result is a local, GPU‑powered speech‑to‑text tool that works everywhere, and I’m amazed it came together so smoothly.

I opened the CLI, stepped away for just 20 minutes, and returned to see a mysterious 10% usage jump. The unexplained increase felt off and frustrating, leaving me wondering why the tool logged activity when I was idle. It’s a confusing bug that made the experience feel unreliable.

I spent an evening tweaking a UI with Claude, only planning to use a small chunk of my weekly tokens. After using about 75%, I expected to be fine until the count jumped to 99% out of nowhere. I asked Dispatch to disconnect everything and tried to understand what happened. The sudden token drain was irritating and left me uneasy about the tool's reliability.

I tried Claude Code to tackle a pull request for Chroma, a Hugo syntax highlighter I barely understood. The AI cranked out patches that got approved and merged, which felt like a win, but I also felt like I was just tossing code over the wall to a maintainer. The tool was useful enough to produce a real contribution, yet my uncertainty about the underlying code left the experience feeling both empowering and oddly unsettling.

I spent hours watching Claude loop the line “All work and no play makes Jack a dull boy,” over and over. It felt like a scene from The Shining, with the AI stuck in a haunting repetition. The experience was maddening and unhelpful, turning what should have been a simple interaction into a frustrating ordeal.

I typed a two‑word prompt, “Commit message,” expecting a straightforward commit suggestion, but Claude ate up 6% of my session and 1% of my weekly quota. The absurd token drain felt wasteful and irritating, so I abandoned Claude Code and reverted to Codex while the issue gets sorted.