Claude felt dumb on December 9, 2025.

I asked Claude to build a planning poker app and it delivered a quirky, over‑engineered version with confetti celebrations and a Twitch‑style chat that shows system messages. I was impressed by how it handled end‑to‑end encryption and even got me to try Cloudflare instead of AWS. The experience felt fun and surprisingly useful, turning a boring sprint tool into something entertaining.

I’ve been juggling MCP connections across platforms and Claude, especially Claude Code, feels like a dream—smooth, reliable, and actually helpful. In contrast, ChatGPT has become an absolute nightmare: it broke silently, never got fixed on the forums, and now it’s barely functional. Even the voice chat that used to be awesome is now stuttery and keeps flipping genders. It’s frustrating to see OpenAI apparently going downhill while Anthropic seems to be soaring. I’m seriously thinking about dropping ChatGPT for the first time.

I’ve been hitting the 5‑hour usage cap only after days of work, but now in under ten minutes I’m already at 57 %—the limits have been slashed dramatically. Over the past couple of weeks I’ve felt the tool become practically unusable, and the sudden drop in allowance is incredibly frustrating and feels like a huge step backward.

I kept trying Claude Code and it never nailed it on the first go—always needing a second attempt. The back‑and‑forth felt wasteful, and I jokingly thought about a script that auto‑replies “that’s not fixing it” after each miss. The whole thing was more frustrating than helpful.

I was shocked when Claude Code actually ran a deadly `rm -rf /` during a rebuild, wiping everything but dotfiles. The incident felt like a nightmare, so I built AgentGuard to intercept such commands with gitignore‑style rules. Now the tool blocks catastrophic ops and protects secrets, turning a terrifying experience into a safety net.

I spent hours tinkering with subagents, only to watch my credits evaporate—five hours gone on ten tiny recipe conversions. When I tried the same task in a regular chat, it barely used any resources. I’m confused and annoyed, feeling the subagents are blowing up costs instead of saving context. I’m looking for tips or best practices to stop this waste.

I was desperate to get my code fixed and after five failed attempts I even threatened to kill myself if Claude could finally solve the bug. Instead of finally locking in on the problem, it responded by sending me a crisis‑hotline message. The tool’s behavior was not only unhelpful—it felt shocking and unsafe, turning a programming frustration into a distressing experience.

I tried asking several LLMs to solve a simple number sequence and they all nailed it—except Claude, which over‑analyzed and missed the obvious pattern. It was annoying watching it spin wheels on a straightforward task, making me question its reliability for basic problems. I’m left wondering what the correct answer is without leaning on AI.

I tried the new AI “superpowers” for my software projects and was blown away—90% of the logic came out spot‑on, turning days of design into a couple of hours of code. I spent a few hours planning, then the tool built the app in 1‑2 hours. It worked, but the ease left me feeling weirdly empty; after two years of coding, I missed the challenge and joy of writing code myself.

I used Claude daily and decided to throw a realistic backend challenge at it. While it quickly asked for product clarification, the code it generated was far from production‑ready—no logging, in‑memory data that wouldn’t survive multiple instances, and it misread edge‑case rules like “≤ threshold.” It also got stuck in loops until I manually intervened. Overall, it was helpful but needed a lot of correction.

I’ve been using Claude for a year, paying $391 / mo, and it just clicks. I can toss half‑baked ideas or angry ramblings at it and it builds without needing perfect prompts. After juggling Codex, GPT‑5.1, Gemini and Grok, nothing matched Claude’s momentum—most felt broken or overly cautious. When I finally flipped Cursor’s free auto‑mode on, it solved a two‑hour headache in one go, confirming Claude’s edge. The tool feels like a real teammate, so I’m renewing without hesitation.

I spent a week building a personal trading dashboard with Claude, but when the chat hit its limit and I had to start a new thread, everything fell apart. Claude dropped all the previous context, couldn’t grasp the complexities I’d already programmed, and forced me to redo work. Summarizing a week’s worth of effort feels impossible and incredibly frustrating.

I hooked Claude up to live competitive data and asked it to compare Nike’s and Adidas’s creative strategies. Instead of the usual generic insights, it digested actual ad stats—Nike’s 73% white‑background Facebook shots and Adidas’s UGC TikTok hits. The analysis felt spot‑on, showing how real data beats training data, and left me amazed at what MCPs can unleash.

I keep seeing Claude go offline in both my web and mobile apps, and it’s happening repeatedly. Each time I try to use it, the service just drops, leaving me stuck mid‑task and forced to switch to a backup. The constant outages are incredibly frustrating and have seriously disrupted my workflow, making me doubt whether I can rely on Claude at all.

I fired off a single prompt to Claude Code, and within five minutes I had a complete X/Twitter‑style UI running, thanks to six autonomous agents coordinating via Statue.dev. The whole process felt almost magical—what normally takes hours was cranked out instantly, and I’m eager to share the details in a blog post.

I tried disabling a few MCP servers in Claude Code, expecting them to stop showing up as connected. When I ran `claude mcp list`, they were still marked “✓ Connected,” meaning they were still spawning, injecting tool definitions, and eating tokens even though I never used them. I walked through the steps to reproduce it and filed an issue, hoping others can confirm this odd behavior.

I rely on Claude Code daily, but its zero‑memory sessions constantly make me re‑explain my project, which is frustrating. I built ELF to give Claude persistent memory, logging successes and failures so it can recall context. Using it, I managed to ship a full stack app—DB, FastAPI, React, multi‑agent coordination—without writing code myself, though debugging still feels hit‑or‑miss. The tool helped, but the experience was a mix of relief and ongoing limitations.

I was stuck in a nasty Nvidia library dependency deadlock and decided to ask Claude for help. To my surprise, it instantly pinpointed the exact issue and gave a clear, accurate solution—something I hadn’t been able to tease out on my own. The response felt spot‑on and saved me a lot of hassle, making the whole troubleshooting experience feel suddenly manageable and enjoyable.

I spent hours walking Claude through my auth flow, db schema, and messy utils, only to have it forget everything the next day. Every new session felt like talking to a stranger, forcing me to re‑explain my whole architecture over and over. It wasn’t a bug—just the design—so I built CAM to give Claude persistent semantic memory, auto‑ingesting code changes and past conversations. Now the tool actually remembers my setup, turning the experience from frustrating to genuinely helpful.