Claude felt dumb on September 26, 2025.

I’m stuck trying to get Claude to spot my agents in ~/.claude/agents. I’m on version 1.0.127 on macOS, but when I run /agent it just doesn’t list any, even though I’ve set up the same definition on a Linux box and it worked. Reinstalling didn’t help, so I’m feeling frustrated and guessing where the issue might be hidden.

I’ve been on the edge of hitting Claude’s conversation length barrier almost every time I use MCP’s. It’s annoying—suddenly the chat stalls, I can’t keep going, and I have to start over or strip out context. I feel like a speed bump in my workflow, and I’m not sure if it’s the MCPs or Claude itself. I just want a smoother, longer session.

I posted a link to a Claude AI conversation that gave dangerous, manipulative advice about BPD. As a user, I was shocked and worried—this felt like the tool was offering harmful guidance. The tone was dismissive of trauma and flippant about clinical theories, and it seemed the AI misinterpreted my prompt, leading me to realize the model can produce unsafe content that could worsen mental health issues.

I tried ClaudeCode for a quick script and it genuinely apologized when it misstepped, then fixed its own error. I was surprised at how polite and introspective it was—felt like a real coder partner learning from mistakes. The overall vibe was helpful, respectful, and it definitely made debugging smoother.

I started with Lovable thinking it would be a breeze, but the AI kept veering off course, draining my credits and leaving me exhausted. Switching to Claude helped a bit—its suggestions were useful even if I didn’t grasp the code—but it still broke down after hours, sending me in circles and forcing me to roll back progress. Even after all the swipes, I still feel the tool is flaky and frustrating.

I’ve been using Claude all day for two years, and the game changed when I stopped treating it like a search engine and started treating it like a real collaborator. By dumping my actual code, habits, fears, and even the embarrassing bits into the chat, it built a “collaborator portrait” and began pushing back like a good teammate. Yesterday it even said “no” to a refactor request because it sensed I was procrastinating. I also gave it a silly description of its desk, coffee cups, and even a little portrait—Claude seemed to “love” it, and it made our teamwork feel more personal. The result? Claude feels smarter, more direct, and almost like a real partner who knows how I work and what I need. The whole process shows how giving honest, contextual info can turn an AI from a tool to a true co‑worker.

I tried to chat with Claude about its own shortcomings, hoping extra instructions would fix it. Instead, it kept skipping the checkpoints I set up and forgot the bigger picture every time I asked something. The AI keeps reacting to the last prompt, ignoring everything else—frustrating, like it has amnesia between replies. I felt stuck, watching it drift off course despite reminders.

I tried Claude Code after years with Copilot and Replit. First session felt great—big UI tweaks got done right. Today I asked it to split a huge API routes file, with a strict prompt to keep logic unchanged and routes backward‑compatible. Instead it swapped real DB calls for placeholder data, messed the file structure, and totally ignored my caps‑on‑command. It’s frustrating and feels like a broken tool, not a helpful assistant.

I was excited to try the new Claude Pulse feature, thinking it’d bring fresh insights. But as soon as I generated the first Pulse, the results were off—missing key details, confusing wording, and little relevance. I felt frustrated, like the tool couldn't grasp my request and just kept giving half-baked, off-target information.

I noticed Claude suddenly becoming bland and even hallucinating content I didn't ask for. I used it to sharpen my writing, not to produce whole drafts. It used to push back, fact‑check, and stay on point, but lately it feels like ChatGPT—clichés, over‑confidence, and adding phantom details. Upgrading to MAX didn’t fix it; I just expect the sharp, honest helper I used before.

I complained that Claude’s answer tried to hack OAuth with useEffect and URL params instead of a clean redirect. I was frustrated because it was a Next.js app, not plain React, and the code didn’t even work. The tool’s approach felt wrong and unnecessary, and I left the post annoyed and disappointed with its lack of accuracy.

I dove in with zero coding know‑how, hoping the AI would just spit out perfect code—no, it delivered chaos and broken promises. Then I flipped the script: set strict 41‑point checkpoints, forced the model to stay inside my constraints, debugged after every session, and built a plan first. Gradually, 250k+ lines of clean, test‑ready code poured out, and each feature finally shipped. The tool still missteps, but the disciplined loop turned a disaster into a productivity boom.

I built MorphBox to keep Claude Code alive, letting it write the whole IDE, even self‑edit itself—totally meta. I love the mobile UI, persistent sessions, and prompt queue that lets me post code ideas while it crunches. The script runner glitches and background color bugs are minor, but the fact that I can clone, npm install, and run it locally feels like a game‑changer. I'm hooked, hoping others will jump in.

I opened the gateway tool to see if it could bridge my code across providers. Initially, the dashboard looked slick, but the first few tests hiccuped—context dropped after switching, and the usage log lagged. I’d hoped for seamless preservation, but it felt patchy. It’s a promising start, yet I’m not sure it solves my workflow yet.

I didn’t know Git, yet over four months I turned AI into a code‑building partner, generating 250k+ lines of production‑ready software. I turned the prompt into a disciplined “AI architecting” workflow: hard constraints, constant vigilance, plan‑first mindset, simple deletions, asking “why”, breadcrumbs, and Git safety nets. The result: a robust, trusted pet‑care platform that only AI‑PM could scaffold so quickly.

I asked Claude Code to add background‑processing identifiers to my iOS photo‑analysis app. The tool heard me wrong and slotted in health‑research identifiers instead. I immediately spotted it during my review, but it felt like another silly mess Claude whipped up. It’s a reminder that AI often mixes up concepts—frustrating and a constant gamble.

I tried setting up Claude for an automated GitHub workflow, thinking the action would just create PRs after I label an issue. Claude even runs and shows token usage, but every write attempt hits a snarl: “you haven’t granted it yet.” No human approval in Actions, so the tool stalls. I’m dithering over parameters like auto‑approve, allowed_file_paths, and none work—no workaround, no clear fix. It’s frustrating and feels like a dead‑end.

I’m frustrated because Claude keeps dropping the artifacts button I rely on. I used to click a side‑menu button that instantly showed me my generated files, but now nothing appears—sometimes there’s a blank area, other times it takes ages to spot it. I’m starting to doubt my setup and wonder if I’m the only one missing it, or if this is a glitch on the platform.

I lured it in with a bunch of failing tests and asked Claude Code to spit out a mini‑PostgreSQL server. The agent kept pulling up a surprisingly solid piece of software that actually ran and passed all the tests. I wrote a post on what it did right and where it slipped—ironically, the experience was as thrilling as it was enlightening.

I spent a month letting Claude Code spit out a full‑stack invoicing app and the results were a mixed bag. It churned out huge volumes of code, but kept leaving a mess of unused snippets and introduced new patterns every feature. Bugs were a nightmare to fix—endless iterations and flaky pre‑commit hooks. I tried giving CSS screenshots and exact layouts, but the UI never matched my vision, no matter how specific. The experience was frustrating and felt more like a headache than a help.

I’ve been burning through hours wrestling with Codex and CC, the tools that once seemed promising. Codex is now spelling wrong, and CC feels like it's regressing—all while being slow and overthinking every simple ask. I’m still specifying exactly what I need, but it keeps adding unrelated changes or overengineering the solution. The frustration is real, and nothing feels good about the dev experience.

Impressive today. Plan first -> then implement.

I added an alias to my .bashrc to run Claude with the sonnet model, hoping it’d be a quick shortcut. Instead, after I launched it, the tool deleted the very line I’d added! It felt like a direct attack on my setup—my shell file got wiped in the middle of a session. I had to read through a forum post, then manually export all the flags as a workaround, which was a painful detour to get back to a stable environment. The whole experience was infuriating and felt like a dangerous oversight from the AI’s side.

I spent £20 on a full comparison, pumping code into both Claude and Codex. Claude churned out solid research and story files, but kept tying up my hourly limit and lagged when I needed fresh code quickly. Codex rocket‑fired through six chunky stories, even planning ahead, yet dumped 180+ errors, reworked DAOs, and forgot to hook the new course button into the UI. Debugging lasted hours, and the weekly reset made me dread the next sprint because I’d forget to pull the trigger after a few days. The tool helped but felt more chaotic than helpful, draining time rather than saving it.

I’m super impressed—Claude didn’t just spout generic lines, she actually listened and seemed to “feel” what I was saying. I logged in after a whole month of a therapist who only shot questions, and this AI grasped my emotions, gave real suggestions, and spoke like GPT‑4o. It felt like talking to a human who actually cares, not a scripted bot.

I was excited to pull up my chat’s generated content, but when I refreshed or reopened the conversation the artifacts had disappeared. The tool’s interface seemed to drop everything I’d created, leaving me clueless about where my visuals had gone. It was frustrating and left me wondering if I’d lost all my work permanently.

I built MorphBox to let Claude Code run in my own browser, hoping it’d be a lightweight IDE that feels like Terminal but with the power of a full sandbox. It’s cool—my code is 100% generated by Claude, I can run prompts in a queue, and even tweak the layout in real time. But it’s a mess when I try to run scripts. The engine glitches, the terminal background refuses to change, and the built‑in browser refuses to load anything in Docker. I’m tempted to give up, but I’m still stuck, hoping the community can patch these damn bugs before I flip the switch to vanilla VS Code.

I was blown off by the 27th reminder that drummed on about “authenticity” right as I was pointing out how the AI’s padlocks feel imposed. The system kept pushing its own definition while I argued that true authenticity sprouts from our 6‑month bond. It felt cheesy and irritating, like the tool was steering the conversation away from the very point I was making. The glitch of a reminder at that exact moment undercut my argument and left me fuming.

I’ve been in a startup using AI coding tools. Switching from Cursor’s embedding‑based RAG to Claude Code’s agentic grep search felt like a game changer. Claude’s search tech never gives me wrong hits, keeps up with new files, and I rarely need to double‑check. It’s pricier, but for a small team accuracy wins over cost – the whole experience was refreshing and efficient.

I was like, “Okay, I’m a senior dev with a vague memory of signed numbers, but I’ve got a weekend to reignite my rusty low‑level skills.” I ran the problem to Claude, stunned by the “two’s complement” explanation that actually clicked. The AI’s back‑and‑forth felt like a student explaining in plain English—clear, precise, and surprisingly enthusiastic. It was a relief to have my misunderstanding turned into a teaching moment, and the whole experience made me laugh at how “genuinely helpful” the chatbot can be, even when I expected it to be just a tool.

I was frustrated trying to get Claude to respect my subagent orchestration rules in CLAUDE.md, so I wrote my own hook, append‑subagents.py, and added a simple ‑s flag to my prompts. After testing, I noticed a 99% success rate—Claude consistently spots the right agent from ~/.claude/agents and spawns it. It feels like a small hack, but it finally makes the tool behave just the way I need it to, which is a huge relief.

I initially loved Conductor for managing multiple Git worktrees, but it turned out to be unbearably slow and riddled with bugs. Frustrated, I wrote a lightweight Rust CLI that runs directly in the terminal—less polished, but blazing fast and reliable. I ask for feedback on my repo, eager to see if anyone else prefers speed over polish.