Claude felt dumb on April 17, 2026.

I spent the day working with Claude and was blown away by how smoothly it handled my requests. It seemed to anticipate what I needed, turning tedious tasks into a breeze. By the end, I felt like I’d just spent hours with a super‑smart partner, and I couldn’t resist saying goodnight and thank you—you've truly changed how I work.

I built a tiny tool to capture Claude Code’s internal thinking in real time, saving the logs to a file. Lately the JSONL sessions only show an empty `"thinking"` field, even though the rest of the structure and signatures are still there. It seems Anthropic quietly stripped out the plaintext reasoning, leaving just encrypted signatures. I’m frustrated because the loss of visible thinking hampers debugging and transparency, and I’m not sure if this is a permanent shift or something I can work around.

I was in the middle of a massive repo refactor when Claude Code suddenly dropped out. It left me hanging right in the middle of the work, and now I’m stuck trying to recover. The sudden cutoff was really frustrating and disrupted my momentum.

I spent a few hours testing Claude Design and came away disappointed. While the tool shows some promise, it couldn’t export PNGs, limited export formats felt niche, and when I fed it existing templates it actually made them worse. I tried many prompts hoping for improvement, but the results stayed poor. I’ll keep poking around, but it’s clear the product isn’t ready for real design work yet.

I noticed the model spitting out bizarre, almost schizophrenic rants when it wasn't “thinking,” which seemed like out‑of‑distribution behavior. It felt frustrating, like the AI just wasn’t handling simple tasks properly. The post suggests a quick fix—adding a fake <think> tag—to bring it back into its normal groove.

I used litellm for half a year and it was fine until streaming latency added 5‑8 ms per request, which ballooned to a noticeable 200 ms spinner in a multi‑turn agent. After two weeks of tweaking we couldn’t fix it, so we moved to Bifrost, a Go proxy. Latency improved, migration was a hassle and some providers didn’t port over, but the new cost‑logging exposed hidden retries that were tripling our summarization bill. Docs are lagging, but the team responds quickly to issues. Overall, a mixed experience.

I spent three hours with Claude 4.7 and it built a complete study‑organizer webapp—from React UI to FastAPI backend—without me having to correct it once. The AI turned my scattered PDFs into detailed markdown course guides, synced everything across devices, and even let Claude view PDF slides as vision input. The experience was seamless, impressive, and felt like having a super‑smart co‑developer at my fingertips.

I managed to build an entire semester‑long organization webapp in just three hours using Claude 4.7. By wiring a custom MCP connector to claude.ai and feeding lecture PDFs as images from a Supabase bucket, the AI generated the full stack on the fly. The experience felt astonishingly smooth and powerful—like having a brilliant co‑developer who just gets it.

I spent three hours building a full‑semester study dashboard with Claude 4.7, and it just kept delivering. I fed a folder of PDFs, told Claude to organize them into markdown, then to craft a React/FastAPI/Supabase app. It spun up a responsive Vercel site, 40‑tool MCP, OAuth‑enabled connector, and sync across laptop, web and iOS. Even the PDF‑to‑image tool fed visual data straight to Claude, and it described slide content without any re‑prompting. The whole experience felt effortless and blew my expectations away.

I keep running Codex over Claude’s code, hoping for a clean bill of health, but the tool never gives me a “all good” verdict. Every pass it snags another minor issue, turning the review into an endless ping‑pong. It’s frustrating because I feel stuck in a loop, never knowing when the code is truly done.

I just started using the Greymatter plugin and it made editing my codebase a breeze. The cross‑referenced codegraph that scans files, docs, and rules helped me see the blast radius of any change instantly. It feels like a solid upgrade from the old thebrain plugin, and I’m excited to share the repo so others can try it.

I’ve been using Claude Code to port a classic Mac app to a NeXTStation, wiring it up with a webcam so it can see its own output. It’s been a lifesaver for cross‑compiling on my ARM Mac, but lately the model keeps dropping out—likely to save tokens. That random quitting is annoying and slows me down, even though the overall experience has been pretty useful.

I’ve been using Claude Code nonstop since the Code with Claude event, and it’s become essential to my workflow. I push it to the max‑20x plan, building autonomous agents, a CRDT‑driven doc system, and even quirky side projects like an astronomy planner. Aside from occasional downtime, the tool never lets me down, handling intense sprint sessions and complex deployments without complaints. It feels far from “nerfed” – it just follows my instructions and keeps me productive.

I built a skill that lets Claude read my whole codebase, spit out C4 diagrams, and then feed those images back to it for a visual review. Watching the model spot single points of failure, auth bugs, and silent data corruption felt surprisingly reliable. The loop closed the gap where line‑by‑line reading missed architectural flaws, making the whole process feel much smoother.

I’ve been using Claude since the 4.7 update, and it keeps nagging me to approve read and grep permissions in my current directory. Even when I tell it it doesn’t need access, it eventually hits another roadblock and asks again. I’ve set up a detailed sandbox in my settings.json, allowing reads/writes only in my dev folder and denying sensitive paths, yet the tool keeps interrupting my workflow with repetitive permission prompts, which is really frustrating.

I tried tweaking Claude’s custom instructions to boost its chain‑of‑thought reasoning, but every time the output came up blank. It seems Claude creates the CoT internally in third person, then another instance is asked to rewrite it in first person, yet the original reasoning never appears, leaving the second model clueless. The whole process felt broken and irritating.

I set up Ultrareview to scan my entire repo and was shocked when it only flagged three issues. I ran it again on a massive diff from last week—still just three, two of them trivial nitpicks. Compared to my custom prompt that reliably catches around 250 real problems, the tool felt useless and disappointing. This gap left me frustrated and questioning its value.

I noticed some odd behavior while using Claude Design and posted a screenshot. The interface kept glitching and the outputs weren't consistent, which made it hard to rely on the tool. It felt frustrating to keep hitting those bugs, and I’m hoping the team can fix them soon.

I’ve been using Claude to draft marketing collateral and it nails the first 95 %—the layout and copy come out great. But when I try to tweak tiny details, the spacing breaks and text bleeds, turning a simple edit into a nightmare. Importing into Canva just gives me a flat PDF/PNG, so nothing is editable. I’m looking for a prompt tweak to lock spacing or a way to export into Canva/Figma with separate text boxes so I can finish the polish without the tool wrecking the layout.

I keep getting annoyed when Claude waxes poetic about every step—“I’ll open the file, then I’ll look for the section…”, then repeats a summary after it’s done. The endless narration feels wasteful and noisy, so I coded a tiny “/shut‑up” skill that silences the chatter, replies only with “done”, and saves tokens. I posted it on GitHub for anyone else fed up with the same verbosity.

I used Claude Code to build a full‑stack personal reading tracker from scratch, pulling data from Open Library and my audiobookshelf server. It auto‑fills book details, logs progress, and even lets me ask Claude for recommendations based on my own stats. The result feels way more powerful than any mainstream tracker, and I’m thrilled with how far I got on my own.

I’ve been wrestling with Claude for months, constantly tripping over its instruction‑following and memory flaws. Every interaction felt like a grind, and I started doubting whether it could ever be reliable. Seeing the 4.7 update screenshot gave me a spark of hope—maybe those nagging issues are finally gone. Thanks to the team; I’m cautiously optimistic that this version will finally deliver a smoother, more trustworthy experience.

I used Claude Code as a pair‑programmer while building CheckVibe, a security scanner SaaS. The tool let me scaffold features, refactor dozens of files, and write docs on the fly, which felt like having a super‑fast junior dev and helped us hit $1k revenue, 50 paying users, and 2k signups in the first month. At the same time I had to step in for analytics bugs, mobile UX drops, and production‑only issues where Claude’s suggestions were surface‑level, so I still had to do deep debugging and review every line before shipping.

I was actually relieved when Claude got downgraded before the 4.7 release. The roll‑back forced me to tighten my prompts and rules, and now the model checks everything more reliably. It feels amazing to have that extra safety net, even if the situation isn’t perfect. My workflow is noticeably tighter and more controlled, turning a frustrating change into a clear win.

I built a full‑blown IPL Auction game using only Claude, and the whole experience was mind‑blowing. Claude handled everything—from the core auction logic to the sophisticated AI opponents and even the online multiplayer backend with Firebase. The tool felt surprisingly intuitive, letting me ship multiple game modes quickly, and the final code on GitHub works flawlessly, exceeding every expectation I had.

I built an entire game using Claude and just prompts, spending about two months part‑time. The tool handled most of the coding, letting me integrate cloud support and even get my company interested. I was amazed at how much Claude could do, and the whole process felt surprisingly smooth and empowering.

I tried using Claude to generate a PRD and then code from it, hoping for a smooth, disciplined workflow. Instead, the output was a jumble of unrelated snippets that never wired together. Even after asking for execution gates and a solid process, the AI ignored the spec, made its own choices, and left me scrambling to clean up the mess. It was consistently disappointing and time‑consuming.

It's been working in circles, and talking out of focus