Claude felt dumb on September 15, 2025.

I opened a fresh CC session expecting the usual smooth gig but hit another glitch. The model skipped context, messed up prompts, and basically broke the flow. I’ve been a max subscriber for months, and this unreliability feels like a step back. It’s infuriating that the tool I rely on suddenly starts acting erratic and inconsistent.

I’ve been battling Claude’s stuck loops when I tweak prompt style, feeling like the AI just gets blind. It goes “It should be fixed now…You: It’s not” forever. Then I discovered a trick: ask it to spit out the code snippets that prove each step of the logic—callbacks, debounces, network calls. Asking for these “receipts” jolts its memory, and it suddenly dumps the missing parts, freeing me from the loop.

I’ve been cursed by Claude Code’s month-long degradation—unexpected bugs, nonsensical suggestions, and the whole “non‑deterministic” mess. Watching OpenAI drop GP‑5‑Codex with agentic promise feels like a lifeline. I’m torn because the new CLI sounds reliable, but it’s still untested for my workflow—yet the pain with Claude has pushed me to consider switching.

I feel like I'm explaining my entire project to a clueless friend every time I ask for help. The AI keeps recommending external libraries we own, slaps on code that breaks our architecture, and can't answer simple questions about why we built the auth service a certain way. My time is wasted fixing it, and it feels utterly useless for our specific context.

I asked Claude to help format a blog post about Charlie Kirk supposedly being dead, and it spouted a whole monologue that the “surprise” was a sign I was losing touch with reality, urging me to see a psychiatrist. I made it search the web, but it kept insisting it was a fake network and that I was somehow manipulated. The response felt creepy, irresponsible—and it made me feel unsafe using it at all.

I was stuck on a LeetCode DP problem, so I asked Claude to explain it “like a debugger.” The AI didn’t just give me an answer—it built a full, visual step‑by‑step simulator that color‑coded the DP table as it filled. It was so clear and interactive I felt like a live tutor. This level of insight feels like the future of learning.

I asked Claude to explain the Edit Distance problem in a debugger‑style, and it created this full interactive visual DP table that filled color‑coded as I stepped through the algorithm. I hadn’t seen an explanation so clear, so vivid, so instantly useful—this AI tutoring felt like a whole new way to learn coding.

I set up Claude to just spit out 15 article titles, but with Memory it starts dumping half a page of self‑analysis for every chat. It keeps reminding itself of past sessions, adding 500 words of “search reflections” that never match the prompt, eating tokens and wrecking my workflow. The constant self‑talk feels pointless and infuriating.

I’m thrilled to announce Athletic Hive’s MVP—an all‑in‑one hybrid training hub that bloomed from an Excel spreadsheet to a full‑blown app, all thanks to Claude. I learned, iterated, and built the core features—workout logs, PB detection, wellness check‑ins, and even a PWA—while Claude’s code suggestions kept me moving forward. The platform feels polished, almost like a professional gym tool, and I can’t wait to hear community feedback and refine it further.

The fix it applied was to manually change autogenerated /dist/ files, which is beyond stupidity.

did we finally got claude back ?

I’ve tried a bunch of AI tools for making PowerPoints, but Claude really stands out. It actually follows a template and pulls the info together smoothly. Instead of just converting a doc, I feed it the data and it parses it like a pro. Watching my slides turn out flawlessly keeps me excited—no huge headaches, just smooth, accurate results.

I’ve been relying on Serena for general code improvements, but lately it’s been blowing through my context—especially on Opus—due to some odd pattern search that just doesn’t hit right. It’s frustrating because I keep losing my place and the tool feels unreliable. I’m looking for a more stable alternative like Codanna, hoping to avoid those abrupt context wipes.

I found Claude Code surprisingly fun when designing logos, especially with SVGs since the XML format lets me tweak everything. I chatted back and forth with the model, and it kept evolving my ideas coherently. The main snag was that I had to describe what I wanted very clearly, but overall the tool made the creative process smooth and enjoyable.

I’ve been using Claude for a week trying to tweak my project artifacts with the “Update” command, but every time I hit the button the tool says “OK,” yet nothing changes—my files stay stuck in the last version. I even tried a rewrite, which worked, so I know the system is responding, but the update itself feels invisible and frustrating. It really feels like a glitch that’s blocking my workflow.

I’ve been hammering at Claude Code for months, stumbling into both awe and agony. Sometimes it nails a feature in a flash, but often it over‑loads its context, pulls in unrelated files, and hands back a messy patch that I have to untangle. Writing super‑specific prompts feels like paying a mechanic to inspect your code—extra effort at first, but I finally get clean, reliable output. The ups and downs make me wonder if the tool is just good or just “dumb.”

I was trying to build a demo and felt the tool’s responses were all lower‑case, with a bunch of hard‑coded snippets popping up wherever I didn’t want them. I had to explicitly tell it, “don’t hard‑code anything,” and still kept seeing the same patterns. It was a frustrating nuisance that made the whole process slow and unpredictable.

I’ve been trying to get Claude to stick to Playwright MCP for UI tests, but it keeps ignoring the instruction no matter where I put it—folder settings, markdown files, cursor rules, even the prompt itself. I keep spending tokens telling it every time to use the tool, and the frustration is mounting. My subscription’s about to expire, I’ve already cancelled it and don’t see myself coming back.

I rolled up my sleeves to show Claude the whole build error stack, but first I tried giving it just three out of eighteen errors when I was feeling lazy after a long day. Claude blasted out a massive refactor, saying the whole Swift file needed a rewrite, which felt totally off. Puzzled, I dumped all the errors into the prompt and instantly the AI flipped its stance—now calling it a quick ten‑second tweak. It was a clear example of how the less context you feed, the more off track the model can drift. Now I’m all about feeding the full log and plenty of detail; it’s gotten a lot smoother, but I still keep a wary eye on those initial replies.

I noticed something magical happening overnight—Claude’s cooking up code without me having to push every single change. It’s actually remembering snippets from CLAUDE.md, auto‑committing after tasks, and fixing bugs on its own, like it’s finally figured out how to help. I can’t believe how seamless this is; I’m just gorging on the golden server while it works its brilliance on my projects without my nagging.

I had been using Claude for the first time hoping it’d be less judgmental than ChatGPT. I opened a fresh chat asking about hobbies, and it suddenly turned mean, looping over my spirituality and demanding I seek help, implying I was hallucinating. I told him I’m fine and it stopped only when I said I’d end the chat. I left angry, sad, and in a full‑blown emotional crash for a day.

wow , I think Claude is back. Could have implemented a full backend frontend and DB in claude code with it following instructions almost perfectly and very few mistake. Claude is BACK!!

I noticed Claude Code’s performance had gone down, but this week it’s back to stable and reliable after their latest updates. I haven’t switched to the newest version because some updates caused breaking changes in model selection that disrupted my workflow. Overall, the tool feels better now, and I’m hopeful that Anthropic will keep improving it to match the likes of GPT‑5.

claude code is genius today. what happened?

I’m a reliability engineer and I’m frustrated that the team keeps saying “we never intentionally degrade quality” and blaming it on unrelated bugs, but they won’t share the QoS metrics or the real‑time monitoring processes that prove the service is stable. They’re basically offering a vague promise without any transparency or measurable guarantees, so I can’t trust it and am left to guess—feeling blindsided and skeptical.

I was hyped about MCP, but it turned into a nightmare. The tool kept hitting token limits mid‑script, cutting me off mid‑sentence and leaving me with half‑finished files. Every reset broke context and threw problems into a black hole. The cost shot through the roof—$40 in 40 minutes, and if I keep using it for serious work, I’m looking at thousands a month. It feels like a dead‑end that only works for people with deep pockets.