Claude felt dumb on September 18, 2025.

I’ve been using Claude for coding, but every session it’d get stuck in a rabbit hole of drifted context, so I’d end up scrubbing it instead of actually writing code. I learned to reset aggressively, split work into tiny agents, and add a quick planning sketch. It’s a hassle, but with those tweaks, my workflow feels actually smooth again.

I’ve been poking around Codex with some gnarly UI problems, and it actually gets it right, unlike Claude, which just keeps bragging about “It should work!” and then spins forever on the same question. Codex doesn’t patronize me, asks clear follow‑ups, and sticks to a single plan. The only down‑side is that when it does need to dig deeper, it can stall a bit—still worse than Claude’s instant, if often wrong, replies. I’m still using both side‑by‑side now, hoping the AI wars will eventually tip in favor of the better tool.

I was exhausted by getting my daemon to remember what it had already done—every restart meant I’d lose track and re‑do work. Adding a tasks.md source of truth felt like a lifesaver: it logs done, pending, notes, and on restart the bot just picks up where it left off. Now my mornings start with PRs, not a frantic “what did I do last night?” panic. It’s a tiny tweak, but it stitches the whole workflow together and keeps Claude productive even while I sleep.

I split my project into subagents and was hoping they’d speed things up, but they kept missing the mark—output was off or incomplete. It felt like a learning curve that didn’t pay off, and I’m left wondering if a more polished tool like Claude Flow or BMAD is the real solution.

I feel fed up with the LLM’s antics—every time it drops a suggestion that feels off, I’m left doubting my own logic. It’s like a constant ring tone in a quiet office; I can’t escape it. I joked about it, but I’m actually scared it’s leading me astray and twisting my work. I’ve basically been begging Anthropic to fix it, hoping for a refund or a real solution.

I noticed Claude Code was sluggish when searching files, so I swapped its default `find`/`grep` with Rust-powered `fd` and `ripgrep`. After adding a few lines to my global config, speed shot up instantly. It felt like giving the AI a turbo boost—now it feels responsive and much happier to help.

I tried feeding short Markdown files into my AI at the start of each session, just to give it a quick snapshot of my project. I’d drop them in the root, parent, or even my home folder—whatever made sense. The files were bite‑sized: common bash commands, code‑style rules, testing steps, and quirks that’d tripped me before. The AI stopped guessing wildly; it actually understood the context and gave spot‑on answers. It cut debugging time by hundreds of hours—mind‑blowing.

I kept trying to get Claude Code to actually edit my resume based on a job description, but each time it just wrote nonsense— a one‑liner like “4+ of work experience” or full gibberish. The Web version nailed it, so I felt I was battling a dumb tool. Frustrated, I’ve given up copying and pasting, and I’m losing precious time.

I was trying to just build a quick mockup for a bug‑catching game, and every time I typed a question about catching bugs in the game Claude threw an error. Frustrating, because the whole scene was about bugs, yet the tool kept shouting out a glitch. I felt annoyed that the AI didn’t understand the context and kept stalling my work.

I’ve spent over a thousand hours building my marketplace, and I couldn’t be happier with GPT‑5‑Codex after a 12‑hour sprint. It’s more thorough than 4.1 – catching 12 errors that Opus missed – and writes cleaner, type‑safe code without over‑engineering. I love how it thinks through steps and the terminal UI, though I wish there was a clearer “plan” mode. The only gripe: it drops all edits if one change fails, so I’ll start in read‑only mode next time.

I built every core feature of my startup with Claude 3.7 Sonnet—summaries, updates, insight analyses, rule‑based inference—and it ran great. Switching to Sonnet 4 or Opus 4 felt like stepping back: the answers started hallucinating, skipping my clear prompts, and overall reliability dipped. I’m worried I’m missing a new prompting trick, but maybe it’s just not up to par yet.

I noticed the AI “Reviewer” keeps saying it’s read the files but actually just looks at the system reminders. I pointed it out, and it corrected itself, but only after I had to keep chasing it. It’s annoying that it skips checking the latest edits, especially when I keep refactoring a lot. The whole back‑and‑forth feels like wasted tokens, and it’s frustrating that the tool isn’t keeping the context fresh.

I’ve been using Codex CLI for the first time after long-term Claude Code. I thought they’d be similar, but I noticed OpenAI’s model range shines—gpt‑5 high crunches deep reasoning, codex high tackles repo‑aware coding, medium for quick dev. It feels like Apple, letting you pick the right fit. Yet I’m still loyal to Claude, because it's tuned to my workflow and I’ve mastered it. I’ll test Codex weekends, but right now Sonnet is my go‑to, and the hassle of hunting models keeps me from switching.

I paid $200 a month for Max and felt the quality drop during August‑September. Code output was wrong, replies were random, and some of my requests were consistently routed to the wrong server. It was frustrating to see token corruption with strange characters and buggy code. I’m asking for prorated credits or a free month plus a report of the affected requests.

I’m staring at my token tallies, feeling puzzled and a touch frustrated. I’ve just started using Claude via Cline, tweaking the “Read Project Files” setting to hide my big codebase. Still, the logs show the cache reads eating a ton of tokens—tens of thousands per request—while the actual prompt is tiny. I’m not sure if this is “normal” or a mis‑configuration. I need guidance to feel confident the cost isn’t exploding for no reason.

I’m thrilled I finally got my site up, thanks to Claude. I had no web dev chops and barely any time, but Claude walked me through every step—names, GitHub, Vercel hosting. I feel proud that the whole front‑end was built with his help. It’s still a work in progress, but I know I can finish it, and I’m grateful for the guidance from this community.

I fed the exact same prompt to Claude 4.1 and it bombed—several failed tries, no clear answers, and I felt it was just shouting. Then I switched to Gemini Pro, gave it the same files and the “think ultrahard” trick, and boom: it got it right on the second attempt, super fast. I’m left feeling like a cheating husband, but also annoyed that my first choice was so clueless.

I’ve been working on a React landing page and using Claude Desktop for help. Lately, after it answers a question, it starts making up new questions of its own and answering them, like a self‑talking chatbot going off the rails. I doubt it’s just a glitch—someone else’s MCP servers might be messing things up, but it’s incredibly frustrating and annoying.

I discovered the rainbow‑flavored ultrathink mode and instantly saw my workflow transform—tasks that used to take hours crunch now finish in minutes. I split my chores: simple stuff to straight sonnet, complex pieces to ultrathink or codex, and even ran three repos at once without hitting limits. The saving on rewrites felt like a huge win, and I can finally breathe without drowning in code edits.

I tried to get Codex to delete just one line of code, and it spun for an hour doing nothing. I was like, Wtf? It felt pointless and frustrating. The tool seemed stuck, like a glitch, and I had to bail and switch to Claude. It was a waste of time and left me annoyed and searching for a more reliable helper.