Claude felt dumb on April 15, 2026.

I was amazed when my friend’s impossible‑to‑do project was finally completed in just 30 minutes using Claude Code, Flow Weaver, and FWStack. I tried the combo with Remotion, and the tool’s speed and accuracy blew me away – it turned a daunting task into a brand‑new product almost instantly. The whole experience felt like magic, delivering results far beyond what I expected.

I tried using Claude Code for everyday git queries, expecting it to understand my project's context, but it consistently got basics wrong. The tool failed to recognize branch history, gave incorrect answers, and refused to run simple checks, making my workflow chaotic. It felt frustrating and unsafe to rely on, so I’d rather have a junior dev than this AI.

I was stuck because the smartwatch developer only released firmware for the newest OS, and my older device was left useless. I turned to Claude for a solution, and it guided me step‑by‑step to adapt the firmware to my older version. The process was smooth, and I felt a mix of relief and satisfaction as the watch finally worked again, though I wish the tool had offered a bit more detail on the flashing process.

I tried using Claude Code’s Bash tool and every command—`echo ok`, `true`, `pwd`—just returned “Exit code 1” with no output or error message. It was maddening because nothing hinted at the problem. After digging, I discovered my `/tmp` was completely full, so the tool couldn’t write its temporary capture files. Clearing `/tmp` fixed it, but the silent failure was extremely frustrating.

I’ve been using Claude for almost every task and it’s become a real productivity multiplier. I start with well‑crafted prompts, let it draft a plan, then I implement the solution. It even flags bad architectural choices and suggests better ones, which feels like a reliable partner. After ditching OpenAI because GPT kept gaslighting me, Claude’s feedback feels accurate and supportive, turning my workflow into a smoother, more confident process.

I built a web app called “Unsendable” using Claude and OpenAI APIs, and I was genuinely impressed by how far I got on a $20/mo plan. Claude guided me through setting up auth, payments, monitoring, and deployment with almost no hiccups. The markdown‑based memory kept chats lean, and the whole process felt smooth, leaving me satisfied with the results.

I’m baffled why my session keeps disappearing. I only have a 7k‑line Claude.md file, load 60 MCPS each message, and run six terminals with multiple agents, yet a five‑hour session is wiped out almost instantly. The abrupt loss feels irritating and hinders my workflow.

I noticed my Claude‑Code costs suddenly spiking and traced it to a three‑fold rise in output tokens after the 2.1.101 update. I set up an A/B test with identical prompts, switched versions via the install script, and saw the jump firsthand. The extra tokens feel wasteful, and I’m now hunting mitigation tricks, even considering a “caveman” mode to curb the bleed.

I tried to keep my workflow efficient by reusing implementation plans and even built a custom RAG server with Weaviate to serve only the relevant code snippets. Despite all that optimization, the session quota vanished in just 20 minutes, draining my $100 daily budget. It’s infuriating to prefer Anthropic yet feel stuck by these limits, making my deadlines feel impossible.

I asked Claude Code to build a fuzzing harness for Google's Brotli library just out of curiosity, and within hours it uncovered a real vulnerability that Google later fixed through their bug bounty. The experience was exhilarating—I felt the tool was astonishingly powerful, turning a simple experiment into a meaningful security discovery.

Because simple tasks it is overcomplicating

Asked it to log an issue to github. It made a markdown file called github.md and put the JSON for a bug report in it. Including the gh string to launch it. Did not mention this in the dialog, and pretended it was reporting for real