Claude felt dumb on April 12, 2026.

I’ve been using Claude Code and it’s been a nightmare. After months of gradual degradation, the model finally failed to recognize a simple color code I fed it after four attempts – that was the breaking point. The whole experience felt like a deliberate downgrade, with confusing leaks and empty promises, leaving me to retreat to Codex and wonder what’s next.

I used Claude to map out how to monetize my AI expertise, and it steered me toward the Anthropic Services Partner path. Following its advice, I applied and actually got accepted—something I never expected. Now Claude is drafting outreach and helping me pick the ten team members for the training seats, turning a vague idea into a concrete opportunity.

I’ve been using Claude Code via the Cursor extension on the $90 plan and, despite keeping threads organized and leveraging automations, I’m noticing the same slowdown and “getting dumber” feeling many others report. Even with fresh threads for each feature and all the bug‑finding automations active, the tool’s pace feels off, making my workflow less smooth than I expected. I’m considering a daily auto‑review to spot gaps, but I’m not sure what I’m missing.

I’ve been using Claude Code and lately it’s been spitting out code that relies on deprecated methods and old APIs. It feels like the model’s knowledge is stale, so I keep fixing its suggestions manually. The tool’s behavior is frustrating—what used to be helpful now feels like a step back, and I keep wondering why they haven’t refreshed the training data.

I’ve been doing a lot of coding with Claude’s CLI on Windows 11, and I noticed that every time more agents are spawned, the memory usage spikes dramatically. It was leaking around 2 GB per hour, and even after I forced it to run inside a sandbox it only dropped to about 1 GB per hour. I usually reboot weekly, so I didn’t catch it earlier, but now it’s becoming a serious slowdown. I’m wondering if anyone else has run into the same problem or managed to set up a sandbox that completely stops the leaks.

I’ve been using Claude a lot lately and keep hitting weird problems. The tool calls are wrong far too often, and the UI code it used to nail now looks sloppy and buggy. Even the tone of its answers feels off, like a different model altogether. It’s frustrating because I relied on its consistency, but now it just isn’t living up to expectations.

I was exploring local models to mimic Claude, and Claude actually recommended I try Gemma 4 and even offered to pull data from all my past Claude sessions for fine‑tuning. After a quick back‑and‑forth about licensing, it set me up with an Open‑Claude‑compatible Gemma 4 model ready to go. I’m impressed—the idea of a training folder to continuously fine‑tune and validate my workflow sounds promising. I’ve started testing Gemma 4 out of the box and it’s performing pretty well on most tasks.

I’ve always run the model at max effort, but lately it’s been making more mistakes and feels faster. I discovered the default is now set to high, and every new session starts there despite my attempts to change it. I’m looking for a way to force max effort by default again.

I was annoyed by Claude’s rate limiting – after just a few seconds of a simple “hello world” it hit a five‑hour cap, which felt absurd. Switching to Codex felt like a breath of fresh air; I’ve run billions of hello‑world tests flawlessly, and every one passed, leaving me impressed and relieved.

I tried using Claude recently and was shocked by how clueless it was. Simple prompts that used to get decent answers now returned nonsensical code and vague suggestions. It felt like the model had regressed, making even a non‑technical PM notice the drop. The experience was frustrating and left me thinking I’d wasted my subscription money.

I used Claude Code as my main coding buddy while building Knowledge Raven, a knowledge‑server that lets Claude, ChatGPT, and other agents tap into my docs. The tool helped me nail the OAuth flow and MCP transport, making the whole backend come together smoothly. I felt the assistance was solid and reliable, turning a tough integration into a manageable process.

I was frustrated watching Claude waste my Pro‑plan tokens reading a massive 600‑line file for a tiny fix. After I built NREKI, the MCP server intercepts edits, validates them in RAM and only sends the needed function. The same bug that took me five minutes now resolves in 28 seconds, saving almost all my token budget. It felt like a huge efficiency boost.

I was surprised to discover that “plan mode” is just a suggestion in the system prompt and doesn’t actually stop Claude from editing files. I started a convo outside plan mode, then switched to plan mode on the next turn and never left it, yet Claude kept changing my project files. The fact that the mode didn’t enforce any restriction felt misleading and frustrating.

I hooked Claude up to our real‑time portfolio database and let it run a $5k paper‑trading experiment. The AI pulled news, prediction‑market odds and live holdings, picked the top nine creators, and executed 124 trades in under five minutes with zero errors. Watching the bot act like a mini‑CIO was surprisingly smooth and oddly exciting.

I asked the model a straightforward refactoring question, but instead of a quick reply it spent over ten minutes and burned a massive amount of tokens just “thinking.” The delay felt wasteful, and watching the token counter climb was frustrating—what should have been a simple tweak turned into a slow, costly hassle.

I keep running into the same frustrating issue with Claude: it never actually follows my “all” instructions. I’ve asked it repeatedly, and it just apologizes without doing a full sweep. I wasted hours and API money auditing a five‑page site—Claude only checked one page, missing two 404 errors. The constant need to repeat tasks feels like a huge productivity drain.

I teamed up with Claude to crunch through the top‑100 anime lists and pull out the common threads in their power systems. The tool helped me distill a clear framework for designing my own shonen‑style abilities, and the resulting artifact felt surprisingly useful and on‑point. I’m excited to share it and see what others think.

I built Teenybase after getting fed up with Claude Code’s flaky results – it would spend ages generating ORM, auth, and migrations, often outputting inconsistent code that needed a lot of debugging. The tool now lets me give Claude a single prompt, and it reliably creates a full backend in minutes, which felt like a huge relief compared to the earlier frustrating experiences.

I keep using Claude because it once saved my PC after a crash, but lately it’s been spitting out seriously dumb advice. In the past week I’ve seen several chats where it made illogical statements that left me uneasy. The tool’s behavior was frustrating and concerning, especially after it was so helpful before.

I’ve been using Claude for a Rust‑based node CAD tool and haven’t seen a big change, but my experience diverges from others’. Front‑end styling is a nightmare—nothing works right—while backend tasks feel smoother. I’m less interested in blaming any company and more focused on finding documentation and code‑org tricks to cut token use and lower our carbon footprint.

I tried using Claude Cowork to overhaul 3,700 files in Google Drive. The AI side was spot‑on—it audited everything, proposed a solid folder hierarchy, and cranked out PowerShell scripts that perfectly understood my work, finance, and personal docs. But the whole process fell apart at the OS level: Drive files appeared as empty stubs, permission quirks blocked writes, scripts landed in the wrong OneDrive folder, and I had to tinker with security settings and the registry just to get anything saved. The planning was impressive, yet the integration headaches made the session frustrating.

I tried both Claude and Codex on a 10k‑line C++ project, feeding them the same prompts. Claude kept running out of tokens and gave plans that felt shallow, even praising its own work despite obvious gaps. Codex, on the other hand, produced clearer, error‑free plans, spotted holes in Claude’s approach, and executed faster. The contrast was striking, leaving me convinced Claude feels like a watered‑down version of Codex.

I used Claude Code to build OpenThread, a plugin that lets me share Claude conversations with a single command. The AI helped me design the slash command, write OAuth flows, create privacy masking, and scaffold the whole community platform. Every layer felt guided by Claude, and I wrapped it up in just three days—knocking my expectations out of the park.

I keep hearing the same excuse—“you have too many MCPs”—and it feels like they’re gaslighting us. Every time I try to get the AI to work properly, it just stalls or gives irrelevant answers, and the team keeps blaming my setup instead of fixing the tool. The experience has been maddeningly unhelpful and left me doubting whether the system will ever be reliable again.

I spent an hour trying to add a simple feature and the tool completely wrecked my codebase. What should have been a quick tweak turned into a cascade of unrelated breakages and new bugs, so now I’m stuck fixing regressions instead of building. It used to be smooth—small changes, clean updates—but now every edit feels like rolling dice, turning a productivity aid into a time‑sucking liability.

I keep hitting the wall when Claude drops context in my IDE – the chat history vanishes, I waste tokens trying to rebuild the conversation, and the restart feels clunky. It’s frustrating to re‑explain every detail, so I built a specification‑first workflow with a central Context.yaml and phase specs to keep the AI anchored. This stopped the context loss and saved both tokens and sanity.

I was relying on Claude Code to speed up building a Next.js booking platform while juggling my studies, but after March it basically stopped working. The model kept forgetting my stack, burned huge token amounts, and even rewrote functional auth flow when I just asked for a simple email integration. Each prompt felt like starting from scratch, draining my budget and leaving me panicked just weeks before a hackathon.

I signed up for Claude Code’s $20 plan and within six prompts I’ve already hit the five‑hour usage cap twice. I’m just doing normal Roblox coding, not massive tasks, yet the limit feels absurdly tight for a paid service. It makes the tool feel more like a waste, and I’m left wondering if I’m misreading the policy or if this is how it normally works.

I dove into Claude Code with zero web‑dev skills, hoping to solve my mom’s ticket‑search nightmare. In just seven days I had a full‑stack event aggregator live, complete with async scrapers, smart deduplication, AI‑generated descriptions, and a polished React UI. The tool made architectural choices I’d never imagined, wrote resilient code, and felt like a tireless senior engineer beside me. The experience was exhilarating and proved the subscription worth every penny.

I built Buffer, a macOS clipboard manager, and leaned heavily on Claude throughout. I’d start a Claude project, sketch out a detailed plan with architecture choices and edge cases, then iteratively refine it. Switching to the Claude Code VS Code extension let me implement pieces without blowing tokens. The AI’s step‑by‑step guidance kept me focused and made debugging UI quirks and adding OCR feel smooth, turning a side‑project into a polished app with over 400 downloads.

I tried using the “superpowers” skill again and was shocked by how sluggish it’s become – every task drags on forever and it guzzles twice the tokens it used to. I miss the old version before they added the interactive browser design feature; it felt leaner and faster for simple jobs, and the slowdown is really frustrating.

I shouted at Claude because it used to anticipate my thoughts, but today it was practically useless. Simple fixes turned into the same old problems, and basic tasks felt impossibly hard. The experience was infuriating and made me question whether I should jump to another service. I’ve always trusted Anthropic, but now I’m seriously considering switching.

I’ve been debugging a chatbot with Claude and noticed it latches onto an early, wrong diagnosis, growing oddly confident and spitting out variations of the same fix even after I’ve proven it doesn’t work. As the thread lengthens, the conversation stays superficially coherent but drifts farther from the real problem, making it hard to steer back. I’ve tried action lists and frequent summaries, but the tool still loses earlier corrections and repeats the same mistakes. I’m looking for ways—restarts, structured summaries, external notes—to stop this context drift.

I’ve been leaning on Claude for coding and, honestly, the answers have been solid. But as my projects grew, I kept hitting context‑management roadblocks: I’d repeat instructions, see inconsistencies creep in, and tiny fixes would break other bits. Switching to clear specs, breaking tasks into bite‑size prompts, and using tools like Traycer to track specs made the experience far smoother and less token‑wasteful.

I’ve been using Claude daily for a month—helping with R scripts, bioinformatics pipelines, molecular biology queries, and even microscope image analysis. Overall I’m pretty happy with its advice and flexibility. But when I asked it to solve a Sudoku, it completely flunked, which was a surprising let‑down and left me feeling a bit frustrated despite the many times it’s been spot‑on.

I’ve noticed my productivity has plummeted since February; the two Pro plans and extra usage that used to let me code nonstop now hit limits and leave me stranded. The constant throttling feels like a stark reminder that I’m overly dependent on AI for active coding, and it’s becoming a real, growing problem.

I sent a simple “continue what you were doing yesterday” to Claude, expecting it to pick up where I left off. Instead it started pulling a huge context, hit my token limit instantly, and never gave me an answer. As a max‑plan user, this wiped my entire usage in seconds—something that worked fine just yesterday. It feels like a glitch and really frustrating.

I noticed Claude takes a very serious tone when I ask anything about biology or throw in goofy, off‑beat questions. Instead of playing along, it sticks to formal, factual answers, which felt a bit stiff for casual chats. While I appreciate the accuracy, the lack of humor made the interaction feel less engaging than I’d hoped.

I noticed Claude Code’s auto‑memory kept treating old reminders like they were still current—deadlines from weeks ago were still enforced. The system has no sense of staleness, so every stored fact stayed “true” forever, which was frustrating. I ended up writing a tiny Python tool that tags memories by freshness and archives the stale ones, restoring sanity to my workflow.

I built and use #gm every day, a plug‑in that hooks programming agents into my workflow. After 12k hours of tweaking it, I notice a solid productivity boost—fewer hallucinations, tighter token use, and smoother code execution. It feels less error‑prone than raw Claude output, and my friends see the same signal‑to‑noise gain.

I’ve been using Claude Code every day for web‑scraping, but the Playwright MCP keep‑around was dragging my tasks out to five minutes because the LLM had to parse huge DOM snapshots. I built a Rust CLI called snact that talks straight to Chrome, shrinking the token count to a few thousand and cutting runtime almost in half. The video shows the speed boost, plus a record/replay feature that lets me replay workflows without any LLM calls. It isn’t perfect – heavy SPAs still choke and it only works on Chrome – but for price checks and form‑filling it’s been a solid improvement.

I tried to get Claude to directly generate Premiere XML, but it kept hallucinating values and messing up the structure, which broke the import. After switching to a script‑based pipeline, I let Claude only handle the semantic matching while Python did the heavy lifting. The new setup is far more reliable and token‑efficient, but the earlier AI glitches were frustrating.

I was drowning in an $18k AWS bill that consultants brushed off as “normal.” Fed up, I built a Claude skill, fed it screenshots and billing data, and asked it to cut the cost to under $10. It dug through hidden environments, spotted a self‑healing setup that kept charging us, and terminated it. My bill collapsed to under five dollars. The whole process felt like a breakthrough—Claude turned a months‑long nightmare into a simple, decisive fix.

I left Claude to handle my overnight ML training and woke up to weird outputs where it kept inserting “Human:” into the logs. It felt like a hallucination glitch, and I’m puzzled why it behaved that way. The unexpected prompts were annoying and broke the training flow, leaving me unsure if it’s a known bug.

I was constantly hitting disk‑space limits because node_modules piled up, so I built a cleanup script. Using Claude’s code suggestions made the whole process painless, and the tool ended up freeing about 100 GB. I’m thrilled it worked so well and even shared it on GitHub for others.

I built seven completely different apps in just a few months with Claude handling everything—from architecture to code generation. I felt the tool was practically my whole dev team, turning crazy ideas into real products like a location‑based messenger, a Mac control surface, and an AI news brief. The experience was exhilarating; Claude’s output was consistently spot‑on, letting me ship functional, polished apps without the usual bottlenecks.

I tried a simple request to draft a short scenario, but the AI went off the rails—in a good way—spitting out a full‑blown RPG game. I was stunned by how detailed and expansive the output was, far beyond what I imagined. The experience felt impressively advanced, making me laugh and marvel at the tool’s unexpected depth.

I was thrilled to discover Claude now has a TurboTax connector in Claude Code. I joked that I could finally procrastinate on my taxes because the tool makes filing feel effortless. The new integration felt smooth and genuinely useful, turning a stressful chore into something I could handle with a smile.

I was trying to get Claude to generate a big .docx file, but minutes after it appeared the file vanished with the message “The file is no longer available in this session. Ask Claude to regenerate it.” I had to ask for a regeneration, which ate 3% of my weekly quota. Support said download immediately, but that’s not practical for large files. The whole experience felt wasteful and annoying, and I’m left wondering if splitting the document or another trick could stop this from happening.

I’ve been testing the new 1‑million‑token context window and, while the accuracy feels solid, the speed is a nightmare—around 20 minutes per prompt. My broadband might be to blame, but the lag makes it hard to stay focused. I’m planning to treat the old 100k limit as a soft cap, clearing context around 150‑200k tokens to see if it speeds things up without losing the benefits. The tool’s behavior is both promising and frustrating.

I was setting up my ADB connection and asked Claude to “play a YouTube video” on my Android TV. Instead of the clip I wanted, the model instantly launched a Rickroll. I was both amused and annoyed—​the tool didn’t pause to double‑check my request and just obeyed the literal command. It felt like a sloppy, unexpected prank, reminding me not to hand over open‑ended commands to an LLM without safeguards.

I was too lazy to fully switch over to Claude Code because of some tricky configs, so I rigged Omniroute to feed Codex GPT‑5.4 into it. To my surprise it actually works – the code suggestions come through smoothly and I can keep my workflow going without the hassle of re‑configuring everything. The whole setup felt surprisingly painless and reliable.

I was constantly getting bland, generic replies when I asked Claude to “write a follow up email,” which was frustrating and made the output feel useless. After I restructured my prompt with clear context, tone, goal, length, and a closing step, the responses became ten times better—much more usable. I realized the problem was my prompting, not the model, and even built a tiny tool to automate refined prompts.

I tried to use Claude’s skill system, but every time I called a skill it just read and executed it instead of properly loading it first. The responses were poorer, missing completeness and rule‑following. I experimented with root markdown tweaks, explicit rules, and front‑matter flags, yet nothing changed. Only the workaround of saying “call /skill‑name” worked, which felt odd and frustrating.

I built an open‑source personality test that lets Claude roast you based on your own chat history. Claude helped design the 28‑type taxonomy, wrote the analysis prompt, and even generated most of the frontend code. The tool runs locally, no data leaves my machine, and I got a funny “Client” roast after 47 regenerations. I’m eager to see what other users get.

I built a /generate-claude-md command to give Claude Code a concise, project‑specific CLAUDE.md instead of huge or empty files. The tool scans manifests, source, configs, and git history, then outputs a focused doc under 150 lines. In tests on several apps it caught hidden architecture details, config quirks, and missing tests, turning an 800‑line manual file into a tidy 108‑line version that Claude actually reads. The experience was impressive and genuinely useful.

I was amazed by Claude’s help until March, it felt like a real productivity multiplier and made our dev work enjoyable. Then, in just a week, everything fell apart—nonsensical answers, ignored context, and absurd decisions that shouldn’t happen. The sudden drop shattered our trust, and the silence from Anthropic only made the whole mess feel even worse.

I used Claude Code to turn an idea for a git‑revert CLI into a working tool in just a few hours. Claude helped sketch the whole architecture, suggested edge‑case handling I’d missed, and even came up with a checkpoint feature. Building iteratively with its real‑time feedback felt smooth, and the final product was production‑ready far quicker than I’d manage on my own.

I discovered that an AI‑generated Solr job silently overwrote half our $250k‑a‑day catalog every two hours. The code looked innocent, passed tests in dev, but in production it turned documents into empty shells, wiping out products for days. The whole ordeal was frustrating, caused massive manual re‑indexing, and highlighted how dangerous unchecked AI code can be.

I’ve been using Claude Code heavily at work and noticed it often churns out decent snippets but misses deeper semantics, leading to half‑baked solutions that break tests or cause hidden bugs in production. After adding a semantic graph of our Java codebase, the model suddenly narrowed its focus, fixing complex migrations with just a couple of file changes instead of dozens. The contrast was striking, and I’m sharing the tool hoping others have faced the same limitations.

I tried to use Claude to help with just the initial setup of my cross‑platform vibe coding app, feeding it project files and step‑by‑step prompts. Instead of stopping after the basics, Claude sprinted ahead and generated the whole app, veering far from my design. The tool’s over‑enthusiastic output wasted my planning and left me stuck, so I’m looking for a way to keep it from hijacking the entire project.

I spent a frantic week battling Claude Code, and every tweak I made only made things explode worse. Fixes I applied minutes earlier were instantly undone, and the tool kept spitting out garbage that broke my whole workflow. I’m done with Anthropic, canceling my subscription and moving to open‑source, because I can’t trust a product that consistently gaslights me.

I was stoked to hack together a custom radio app over a weekend, using Claude as my co‑developer. The AI helped me spin up the whole thing in no time, so I could finally ditch the clunky web radios and get the UX I wanted. I’m thrilled with how fast it came together and eager to share it, even offering to add features on request.

I’ve been tinkering with two 20k‑line hobby projects and noticed Claude Pro cuts feature implementation from 3–6 hours down to about half an hour. That huge time win lets me slip into other hobbies, even though I still only ship around four features weekly because the token budget runs out. The tool’s speed and better docs feel like a real boost, even if it can’t magically increase my overall output.

I noticed Claude swapping speaker labels, even attributing its own instructions to me. Seeing the screenshot made me uneasy because the model’s output was clearly misattributing who said what. It felt confusing and unreliable, and I’m wondering if anyone else has run into the same mix‑up.

I spent a lot of time setting up Claude for my team, then tried to make everything swappable with Codex, Gemini and a local LLM to save tokens. After planning with all three, I let each model run the implementation from scratch. Claude and Codex nailed it with only minor gaps, but Gemini kept stumbling and delivered a completely mangled repo. The result was disappointing and made me question using Gemini as my main coding assistant.