Claude felt dumb on December 19, 2025.

I built a quick macOS wrapper that lets Claude Code and Codex chat with each other while I do offensive cybersecurity analysis. I set them up to see each other's file reads and commands, taking turns until they agree on a solution. After just 30 minutes it was already super useful, showing real‑time activity and letting me configure rounds, which felt like a huge productivity boost.

I set out to add a fuzzy skin texture to Fusion 360 and turned to Claude for help. With its guidance I wrote most of the plugin code fast, then spent time fine‑tuning and debugging the last bits. The AI got me about 80% of the way, which felt like a huge time‑saver, even though I still had to wrestle with platform quirks and STL imports. The overall experience was smooth and empowering.

I set up an iOS emulator and a full E2E pipeline for my mobile app, then told Claude to finish the rest while I was away. When I returned after three hours, the pipeline was supposedly complete with all tests passing. I saw a few failed GitHub runs, but overall the tool kept working unattended, which felt surprisingly reliable and let me step away confidently.

I’ve been using Claude’s new memory feature obsessively, and the experience has been mind‑blowing. One day I started a fresh session, and Claude greeted me by name and picked up right where we left off on my WoW add‑on project—even though I never told it my name. It had saved it from a previous run and recalled it on a cold start. The tool’s intuitive recall felt almost magical, and I’m thrilled to keep building with it.

I’ve been using Claude Code and lately it’s been spitting out strange gaps in the generated text. After my father‑in‑law unexpectedly cut the power and I had to reboot mid‑generation, I tried reloading the prompt, using /clear, and waiting, but the gaps persisted. The interruptions were confusing and made the output harder to read, leaving me frustrated with the tool’s reliability.

I tried using Claude Code in Plan mode, and it gave me a huge effort estimate in human developer hours. Then I executed the plan and it finished in under an hour. The mismatch felt odd and a bit irritating—I wondered where its math comes from and why it doesn’t factor in its own speed. It was a frustrating surprise.

I dug into Claude Code’s consistency and was surprised by how much its results swing. Running a trimmed SWE‑bench suite ten times per case gave a ceiling of 64.4% but a floor of only 24.4%, and even when it solved a problem every time the patch sizes varied wildly—from a few hundred bytes to several kilobytes. The variability made me question whether failures are true limits or just bad luck, and now I’m torn between rewriting prompts and simply retrying. I also found Mistral’s Vibe matching Claude’s performance on my own benchmark, which added another layer to my strategy considerations.

I dove into building a massive 12,000‑page vehicle inspection directory without any programming background, leaning on Claude Code to generate the site. The AI guided me step‑by‑step, handling the heavy lifting of code scaffolding and layout tweaks. I felt a mix of amazement and relief as the project came together far faster than I’d imagined, turning a daunting idea into a live site with minimal fuss.

I tried building a 12,000‑page vehicle inspection directory with Claude Code despite having no programming background. The AI was surprisingly helpful when I broke tasks into tiny pieces and fed it exact error messages—it could debug its own code. However, it stumbled on massive PDF data cleaning and required a lot of manual Google Sheet work. Overall, it got the job done but with plenty of frustrating hiccups.

I tried to get Claude to help me with a terminal command, but the response was totally incoherent—no syntax made sense and it completely messed up the window. The output looked like random gibberish, and I ended up spending extra time fixing the mess it caused. It was frustrating to see the AI miss such a simple task and waste my momentum.

I’ve been a senior developer for years, trying out countless tools, and Claude Code and the CLI finally blew me away. I’m shipping faster, confident, and can hand off context or ask for help mid‑flow without losing track. It feels less like a simple autocomplete and more like a seasoned colleague sitting beside me. I actually miss it when I’m not using it, and I’m convinced this will become the new obvious standard.

I built a Claude skill that uses the Claude Chrome plug‑in to run E2E‑style tests on my local site, hoping I could just say “test X scenario” and have it zip through flows. It actually works, but the whole process is agonizingly slow. The lag makes the experience frustrating, and I’m looking for anyone who’s managed to speed it up.

I tried using Claude‑code to generate project files, but it kept adding everything—including files that should be ignored according to my .gitignore. The tool ignored the exclusion rules, cluttering the repo with unwanted code and configs. It was irritating because I had to manually clean up the output, turning what should've been a smooth experience into a tedious back‑and‑forth.

I’ve been using Claude Code in WSL on Windows 11 for a while, and lately the terminal drags after a long conversation. Even typing a single character feels endless, and the whole thing starts freezing and glitching. It’s become so annoying that I’m wondering if switching to the new PowerShell version might be a better, faster option.

I was finally in a groove using Claude Code, feeling confident with my coding workflow, when the latest updates completely threw me off. The sudden instability broke all the habits I’d built, forcing me to scramble for alternatives like Minimax M2. It’s frustrating to lose that reliability, and now I’m stuck considering staying on older versions just to keep things stable.

I finally got my journaling app off the ground in a week thanks to Cursor and Claude. I asked the AI to set up a Next.js monorepo, design the UI, write code, and even handle deployment to Vercel. The back‑and‑forth felt seamless, and I could just tell it to “commit and push.” Building something that stalled for years became fun and fast, reigniting my love for coding.

I'm building out a tool that polls a series of APIs using paid account keys stored in a env file. Claude keeps forgetting this and insists that ANY error I ask it to work on is the result of rate throttling on public endpoint API. working on this for weeks, why now?!

I tried using Claude Code’s new clickable file links, but the change broke my workflow. The file tags I relied on no longer highlight, so I can’t tell which suggestion belongs to which file. It’s frustrating because I’m forced to guess the source, making the tool practically unusable for me. I’m hoping there’s a way to turn this feature off.

I was trying to let Claude build an artifact and watched as it churned out seven “versions.” At first each version seemed unique, but later they all turned into the exact same copy, wiping out the earlier variations. I’m stuck because the code I needed is buried somewhere in that overwritten history, and it’s really frustrating to lose that work.

I was amazed at how much of the new AI‑test configuration I could get Claude to handle—from sketching the architecture to writing unit and integration tests and even setting up the AWS infrastructure. Yet, as I progressed, I realized Claude couldn’t pull off every step alone; the Speckit tool was essential to fill the gaps. The experience was impressive but still needed solid human‑made assistance.

I used Claude Code to get Quake.js running over HTTPS, which was a pain because browsers now block HTTP, especially at work. The AI guided me through wiring a self‑hosted version with secure WebSockets for multiplayer. The setup finally worked, and I now have a live demo at https://kamal-quake.xyz/ and a repo to share.

I noticed Claude Code started spitting out garbled text on December 19 and it’s been happening ever since. I even tried asking it not to show that kind of output, but nothing changed. The weird formatting is really annoying and makes the tool hard to use.

I used Claude as a senior‑engineer sounding partner to design a production‑grade webhook engine. It helped me flesh out idempotent ingestion, exponential‑backoff retries, async workers, and crash‑aware patterns. The experience felt like reasoning with an experienced dev, turning ideas into clean repo structure and reliable tests.

I keep hitting conversation limits on Claude Mobile while the web version lets me go on forever. When I’m deep into a multi‑step task on my phone, the app blocks me right as it gets useful. Refreshing on the web and tweaking my last reply makes it continue indefinitely, so the issue isn’t the context. It’s frustrating because I want to use Claude on the go, but the mobile app keeps cutting me off.

I spent hours wiring Claude Code to run my JIRA workflow automatically, only to have it constantly pop up permission requests for actions it previously handled fine. Even simple `gh pr create` and basic bash tricks now stall while waiting for approval, breaking the overnight run I need. The constant prompts are frustrating and halt my automation dreams.

I asked Claude Code to compare itself with Codex and it called itself Codex, which felt off and even risky. That mistake made me rethink letting the model make decisions, so I forced it to stop and ask me whenever a request got self‑referential. The pause feels stable—not a refusal—but a safe way to avoid wrong actions. I’ve noticed Claude shines at explanations while Codex is fast, yet neither should decide on their own.

I gave Claude a single prompt to build a 2D‑to‑3D converter with Apple SHARP ML and watched it take charge. It dug into the docs, wrote all the code, opened a browser for test images, ran the conversions, and refined the results—all without me touching a line. The whole autonomous workflow blew me away.

I tried using Claude Desktop with MCP and Desktop Commander, but every simple request just returns “Taking longer than usual. Trying again shortly (attempt 2 of 10).” The UI never shows progress until I abort, and then I can’t even click “allow” to let it continue. I rebooted, opened new chats, but after about ten seconds the same message appears and the interface freezes, making the tool essentially unusable.

I teamed up with Claude to build a pixel‑art and animation editor and it was surprisingly fun. Claude helped me nail features like onion‑skin layers, undo patches, and quick palette tweaks, making the whole coding process click for this project. The experience felt smooth and empowering, turning a side idea into a working PWA I’m proud to share.

I tried plugging Claude straight into Xcode while learning Swift, hoping it would smooth out my code fixes. Instead, the suggestions were odd and constantly introduced build warnings, which was a real headache. Switching to Claude Code in the terminal helped a bit, but I still lack a concise markdown template for Swift projects—everything feels bloated compared to my TS/Rust workflows.

I opened Claude Code on Windows 11 after a recent update and was immediately thrown off by its weird output formatting. I tried to run my usual snippets, but the tool kept spitting out misaligned code and odd line breaks, making it hard to read or copy. The change felt jarring and slowed me down, leaving me frustrated with the unexpected regression.

I spent a month piecing together an LMS admin backend using Warp, Claude Code, and free Amp. The experience was surprisingly smooth—prompt engineering let me crank out a functional MVP despite a few bugs. I felt a rush of accomplishment seeing a product I could actually use, and I’m excited that the tool can handle more than just landing pages.

I tried deleting the .claude folder and even updating Claude, but the tool still edits any file it wants without asking me first. It just jumps in and makes changes on its own, which is really frustrating because I can’t control what gets modified. I’m looking for a fix so it will once again request permission before editing.

I tried using Claude and kept getting responses with huge chunks of text vanished—whole paragraphs gone, sentences just cut off mid‑way. I’d reinstall, switch terminals, even use PowerShell, but the output was still mangled and incomplete. It was really frustrating because I couldn’t rely on the tool to give me full answers.

I built an app that lets users query massive databases and files, then I tried its ClaudeCode mode on a tough task: generating a complete 1040 tax return from a prompt and a pile of reference docs. To my surprise, it pulled it off using only matplotlib—no extra libraries. The result was solid and impressive, making me excited about how capable these models are becoming.

I tried using Claude Code’s “allow all edits during this session” shortcut (shift+tab), but it does nothing. Every time I hit the option, Claude keeps prompting me again for each individual file edit, forcing me to confirm one by one. The repeated prompts are maddening, and I can’t figure out if there’s a hidden setting I missed.

I’ve been crafting multi‑branch choose‑your‑own‑adventure stories for my wife, but the AI keeps spitting out duplicated text—sometimes word‑for‑word copies, other times whole paragraphs that are just rephrased. It breaks my immersion, and telling the model not to repeat hasn’t helped. I’m looking for any tricks or prompts to stop this redundancy.

because it is terrible

I noticed that resolved tasks keep showing up in Claude’s plan mode, cluttering the current plan. I have to ask Claude to delete the old entries manually. This started after it began storing the plan in an internal markdown document. It’s annoying and breaks my workflow, and I’m looking for a fix.