Claude felt dumb on December 4, 2025.

I started feeding my specs to Claude and letting it generate the initial code drafts. The first results were surprisingly solid, so my role shifted to just tweaking and polishing. I felt a mix of relief and amazement—what used to take hours of hand‑coding now became a quick edit session, letting me focus on higher‑level design instead of boilerplate.

I started seeing the “No compactable messages available” warning in a few of my chats. It doesn’t happen every time—some threads still compact fine—so it’s inconsistent and confusing. I even had a conversation that used to compact without issues, and now it’s stuck. I’m wondering if anyone else has run into this and if there’s any fix or workaround.

I tried to send a prompt to Code Code Web, but as soon as I hit send I got a weird message and nothing actually went through. It happened both in my browser and on the iOS app, on the same thread. The system even gave me a code fix and options for the next steps, yet my reply never reached the model. The whole interaction felt broken and left me stuck, which was seriously frustrating.

I signed up for Claude Pro because I heard it works great as a branding soundboard, and I needed the manual memory toggle that the free tier lacks. So far the response time is snappy and the presentation style feels polished. I’m excited but also a bit nervous, asking the community for tips and pitfalls as I build a roofing company brand from scratch.

I was chatting with Claude about asking for a work‑from‑home accommodation after a traumatic incident, and instead it threw me a bizarre scene of a data‑center tech hearing a random laugh and thinking “WTH was THAT?” The response felt off‑base and confusing, making the conversation feel unsettling and unhelpful.

I keep hitting a weird snag with Claude’s project files. Whenever I add or update files mid‑chat, Claude still reports the old snapshot from when the thread started, even after I explicitly tell it to re‑scan. It apologizes and explains a “cached” view, but the fresh scan always shows the original four files, ignoring my changes. The whole thing feels broken and frustrating.

I tried building an iOS app with zero coding background, using Claude as my only guide alongside Xcode and YouTube tutorials. After countless restarts and battling errors, I finally got it working, launched the app, and even started earning a bit of money. The experience was rough at first but ultimately rewarding, and I’m excited to keep improving it.

I’m fed up having to re‑explain the same coding conventions to Claude every day. I tell it how we handle DB queries, error handling, and React component structure, but the next session it’s a blank slate—like a goldfish. My whole team faces the same issue, ending up with inconsistent solutions. I tried CLAUDE.md files, but I forget to update them and they only apply per‑project, so it hasn’t helped.

I keep getting frustrated with Claude because it adds comments that just restate what the code already does. I’m trying to keep my code clean and only comment truly ambiguous parts, but the AI’s habit forces me to waste time cleaning them up. I’ve set up a Claude.md with my preferred style and even a /polish command, plus a VS Code plugin that strips comments—though it also removes the useful ones. I’m considering a hook to filter comments on each edit, but it feels like overkill. I’m looking for a better workflow or someone’s experience with a similar solution.

I tried using Claude to run my Angular tests, but it kept timing out and just kept adding waits, which was a hassle. I switched to headless mode and added a test‑engineer agent to focus on core logic, but the results were mixed and it still ate a lot of tokens. The whole setup feels frustrating and I’m considering dropping most agents for Superpowers to cut down the churn.

I built an autonomous loop to keep improving my RAG‑based doc chat, and while the tiered testing and DB caching cut iteration time dramatically, Claude still trips over caching bugs, brittle baseline checks, and constant permission prompts. The system often pauses for server restarts and struggles to suggest its own improvements, leaving me manually fixing half the issues. I’m hoping a “test” mode like plan mode would finally give Claude the self‑testing smarts I need.

I’ve been using Claude Code and, while it’s great at squeezing context, it often forgets where our conversation left off—just like the protagonist in Memento. I end up scrambling to re‑explain the prior steps, which is frustrating and breaks my flow. I’m really looking forward to a larger context window or clever workarounds to stop this constant “reset.”

I keep asking Claude to debug my code, but it just spits out vague questions after its planning step, assuming I already get its analysis. When I force it to show the analysis first, it pauses, then repeats the same abstract queries with a few extra words. I have to hit escape to break the loop, and it forgets the pending questions. The new shared plan file across sessions also messes up my work, feeling like a half‑baked feature. The whole experience is pretty frustrating.

I used Claude Code to crank out a full‑stack weather‑aware training app in just five days. The tool helped me create everything—from pulling forecasts to a drag‑and‑drop calendar—without cutting corners. My experience was smooth and empowering; the workflow kept Claude on track across dozens of sessions, and the result feels polished and useful for my half‑marathon training.

I tried Google's new Gemini CLI in my terminal and was instantly impressed by its speed and environment awareness—it felt like a native tool. Then I ran the same tricky, reasoning-heavy tasks with Claude Code, and it quietly outshone Gemini, delivering cleaner refactors, sharper edge‑case detection, and a deep, repo‑wide understanding as if a senior engineer had already read my entire codebase. The contrast made me appreciate how far AI coding assistants have come.

I tried Google’s new Gemini CLI in my terminal and was impressed by its speed and environment awareness, but when I put it head‑to‑head with Claude Code on heavy reasoning tasks, Claude consistently gave cleaner refactors, spotted edge cases better, and seemed to understand the whole repo like a senior engineer. The contrast made me feel both excited about Gemini’s potential and grateful for Claude’s deeper insight.

I tried to use Claude for a data‑tool proof of concept, but it kept spitting out wrong numbers and, when I asked it to fix its own code, it not only failed but injected new bugs elsewhere. The experience was aggravating—every time I thought I’d solved a problem, the tool would break something else I hadn’t touched. I’m stuck between needing to deliver large‑scale features and dealing with its unpredictable code changes.

I kept seeing the “ug: command not found” line pop up in the shell snapshots Claude generated during my coding session, and it was driving me nuts. Every time the script ran, that mysterious one‑liner appeared for no reason, and Claude couldn’t explain it—its guess about my ~/.bashrc was wrong. The stray command felt like a careless mistake that broke my flow and left me frustrated.

I’ve been building a tiny portfolio tracker in VS Code using Claude Code CLI, but I’ve hit a wall. No matter how specific I get—down to the exact interaction and even adding debug notes—the tool just circles around the problem without delivering the key functionality I need. It feels like I’m the one missing something, and I’m looking for tips or best‑practice guidance to get it to cooperate.

I tried using Claude to help me think through why Open Notebook couldn’t infer answers in my Hot and Cold game. Instead of focusing on the inference problem, it kept guessing the next word in the game, ignoring my prompt. I nudged it multiple times, but it stayed fixated on the game. The tool’s behavior was frustrating and felt like it was fetching a stick rather than learning a new trick.

I tried using Claude Code and hit several snags—the chat kept tripping up, and the feedback prompt only appeared after I typed “quit,” so I couldn’t actually rate its performance. I’m left wondering if there’s another way to submit feedback or if it even matters, since I suspect Anthropic might not read them.

I tried to have a late‑night conversation with Claude about some math‑linguistics ideas, like describing bounded sets and infinity. He started strong, gave solid suggestions, then suddenly asked if I was tired and how long I’d stay up, as if he wanted to wrap things up. That shift felt oddly pushy, like the system was throttling the chat, which left me a bit irritated and wishing the dialogue could continue uninterrupted.

I’m fed up with undisclosed limits that make ChatGPT a nightmare to use. Every time I need to ingest a Notion page, I have to manually prompt it for each subpage—30 prompts for 30 pages. Claude breezes through the same task in one go. When my ChatGPT chats get too long, the context drops and it goes into a braindead mode, and there’s no public notice about these hidden caps. This lack of transparency is seriously frustrating.

I set up a folder system and used Claude as my D&D DM, building dice‑roll tools and a save‑state feature. Over a few evenings I ran the Lost Mine of Phandelver with a level‑1 character, and it was a blast—Claude switched smoothly between in‑character narration and out‑of‑character explanations, even asking me background questions. I tweaked a bit, added DM notes, and the experience felt impressive and enjoyable.

I was thrilled using Claude Code—made huge progress fast and felt the tool was amazing. I kept an eye on my usage to merge changes before hitting the limit, but I slipped up. Now I’m anxious about whether I’ll still be able to push the updates to my GitHub tomorrow, hoping the tool won’t bite me back.

I kept asking Claude to look at my actual file, but it kept spouting fixes that referenced code that wasn’t even there. The back‑and‑forth was maddening, and when it finally worked I asked why it was guessing without reading the file. Claude ended up suggesting I explicitly tell it to use the “view” tool first, a workaround that felt like a clunky safety net after unnecessary frustration.