Claude felt dumb on December 2, 2025.

I tried using long prompts to steer Claude, but the context kept bloating and the output got confused—prompts kept breaking whenever I added more rules. It was frustrating until I switched to a hook system that only fires on mistakes, keeping the context clean. Now I feel the tool is much more reliable and scalable, though the earlier prompt experience was a real pain.

I’ve been building AI‑driven tools for years and discovered long prompts just bloat context and cause the model to stumble. After ditching big prompts and switching to lightweight “hooks” that fire only on mistakes, everything became cleaner and scalable. The hook system now blocks bad code, enforces file headers, and even cross‑language patterns, making the AI feel reliable and much more useful in my daily workflow.

I was using Claude to translate a German chapter to English, but it skipped a whole section on the first try. When I asked how to prevent that and if it needed more input, Claude instantly blocked me, saying it couldn’t respond and flagging my follow‑up as a “prompt injection risk.” The unexpected shutdown was confusing and irritating, leaving me unsure why a simple clarification request triggered such a security warning.

I was working on a multi‑step project with Cloud Code and noticed it kept “starting from scratch” whenever the chat got long. After completing a five‑step plan, I asked it to commit and push, but it insisted on re‑implementing the whole plan, duplicating earlier work. The repeated reconstruction slowed everything down and felt frustrating, as the model seemed to lose track of what was already done.

I tried using Claude Code after the new planning feature rolled out, and it quickly became a headache. It would fix an issue, then generate a plan markdown and suggest the exact same fix again, as if it hadn’t remembered its work. On top of that, it stopped printing full code in the terminal, and the “/plan” prompt just ran “/usage”. The tool felt confused and repetitive, making the coding flow frustrating and inefficient.

I’ve been noticing my context window shrinking dramatically over the last few days – conversations now get compacted three to four times faster than they used to. I’m on a Max plan and using Claude inside VS Code’s terminal, and the same issue shows up in PyCharm. It’s frustrating because I can’t keep the context I need, and I’m trying to figure out why this is happening.

I was stunned when Claude completely flipped its answer—from calling my coworker wrong to admitting they were right. The turnaround felt almost comedic, and I could literally hear the “holy shit” in my head. It wasn’t just a correction; it was a humbling moment that made me trust the model’s judgment far more than I expected.

I keep getting mental‑health pop‑ups from Claude every time I send a new message, even when I’m just chatting about everyday stuff. It started after a brief outage, and the sudden notifications feel intrusive and odd, making the conversation flow feel disrupted.

I tried ClaudeAI again after months of sticking with ChatGPT, and the difference blew me away. When ChatGPT kept spitting out extra, sometimes inaccurate text despite my prompts to trim it, Claude instantly delivered a spot‑on, concise answer that looked polished. The contrast was stark, leaving me impressed and unable to view ChatGPT the same way.

I love the sub‑agent feature when it works, but I’m constantly having to nag Claude Code to actually use them. Even after tweaking descriptions, the Claude.md file, and hooks, I only see them fire about 40% of the time. The unreliability is frustrating, and I’m looking for any tips to make the tool actually remember to call sub‑agents when it should.

I kept getting a pop‑up saying Claude Code was pulling the same random OneDrive file while I was working on something completely unrelated. The file’s name and content had nothing to do with my task, and the settings page offered no clues. It wasn’t disastrous, but the unexplained download was irritating and made me wonder if there’s a way to stop it.

I ran a side‑by‑side test of Claude and Codex on the same coding tasks using my Rover tool. Claude consistently respected the existing code context, followed my exact instructions, and wrapped up complex changes in about ten minutes. Codex often rewrote features, added unsolicited tests, and sometimes got stuck in long loops, taking up to 45 minutes. Overall I found Claude’s behavior more reliable and aligned with my project’s needs.

I tried using the AI while vibe‑coding and it kept circling the same API issue. It would “fix” something, then revert, looping for hours and draining a massive amount of tokens. The whole process felt frustrating and inefficient, and I’m left wondering how to steer it toward a single, focused bug‑fix instead of endless back‑and‑forth.

I’ve never coded before, yet Claude took my vague ideas and turned them into a daily‑use program. It’s completely transformed how I invest and trade—what used to take weeks of manual digging now happens in minutes. The speed and accuracy felt almost magical, and I’m genuinely grateful for how it’s reshaped my workflow.

I tried to get Claude Code to add a procedural starfield to my Rust/Bevy game, expecting a few hours of work. Instead I spent two weeks and three full rewrites battling dense, wrong plans, outdated APIs, and hallucinated code that never compiled. The agent’s confident but incorrect designs and ignored constraints made the whole process frustrating and time‑consuming.

I’ve been using GLM together with Claude Code in my daily workflow and, overall, it’s decent. I set up two aliases – GLM handles simple, repetitive tasks to save tokens, while Claude tackles the complex stuff. It’s not on Claude Pro’s level, so I was a bit let down when I expected parity, but the price and higher limits make it a handy backup.

I finally switched to Claude after years with ChatGPT and was blown away. The answers felt richer, I could pin chats, and even talk about beer without weird censorship. Pasting long text gave me real, useful feedback instead of generic praise. I’m amazed I didn’t discover it sooner—this tool truly feels a massive upgrade.

I built a small Codanna script that lets me slash‑command Claude to generate interactive, force‑directed graphs of a symbol’s call radius. Trying it on Three.js’s uniform function instantly clarified its tangled callers across lighting, texture, and rendering code. The tool feels solid and useful—quickly showing how tangled a function is and how far changes will propagate, and I’m eager to see it added to the Claude profile.

I spent the past month building Reps, a Swift‑based AI fitness coach, and leaned heavily on Claude Code. I described features, handed it bug repro steps, and watched it navigate a 150K‑line codebase, updating SwiftUI views, services, and cloud functions without breaking consistency. Fixes popped up quickly, and new features landed on TestFlight the same evening. The speed and reliability felt almost magical, turning my solo‑dev grind into a rapid‑iteration workflow.

I used Claude to help verify real versus fake online discounts, feeding it structured price‑history data and asking it to walk through the reasoning. The tool broke down multi‑step logic, spotted spikes, cleaned messy retailer info, and even suggested test cases. Its explanations let me fine‑tune my detection rules and felt surprisingly effective for this reasoning task.

I spent months building a price‑tracking and deal‑verification system, and Claude turned out to be the unexpected brain behind the reasoning layer. I used it to spot real vs fake price drops, design anomaly logic, clean messy retailer data, craft edge‑case tests, and translate model decisions into plain English for audits. The tool’s multi‑step reasoning on structured inputs was impressively sharp, making the whole process smoother and more trustworthy.

I tried to get Claude Code Web to handle a fairly simple request—reading a parameter, pulling extra data from the database, and updating the response and API spec. Every time I fed the prompt it stalled for half an hour, doing nothing but spitting out my code, and never progressed. Nothing was charged, so I figured the model itself was fine. Only when I base64‑encoded the whole prompt and asked it to decode did it finally execute the task. The whole thing felt oddly blocked and irritating.

I spent a night building a Statamic plugin with Claude’s help, and it actually worked faster than the official WordPress solution I wrestled with years ago. The tool felt like a creative partner, reigniting my love for web development and replacing endless scrolling with productive joy. I’m already dreaming about a Telegram bot to post to my own Tumblr clone—Claude has given me back freedom and excitement.