Claude felt smart on March 21, 2026.

I dove into building an app with zero coding chops, just talking to Claude Code. It wrote all the Swift files, set up APIs, and guided me through terminals, databases, and deployment. I felt a mix of excitement as ideas materialized and frustration during endless polishing, but ultimately the tool turned my vague concept into a ship‑ready product and reshaped how I handle information.

I’ve gotten pretty good at coaxing Claude Co‑worker into generating full design systems—components, typography, colors, icons, the works, even a skill.md to avoid “AI slop.” But when I switch to the actual build phase with Claude Chat, the layouts turn out terrible despite using the same detailed prompts. The loss of context kills the quality, and I’m stuck wondering how to transfer the whole design‑system mental model into the build chat—maybe design tokens? I need a better way to keep the knowledge intact.

I’ve been using Claude Cowork to draft complete design systems—components, typography, colors, icons, everything—and I usually get spot‑on results. But when I switch to the actual build phase with CC, the layouts turn out terrible despite using the same detailed prompts. The loss of context is wrecking the output, and I’m stuck looking for a way to transfer the whole UI model into the build stage, maybe via design tokens or another method.

I was fed up with Claude Code constantly guessing file and table names, breaking my healthcare SaaS and wasting half an hour per bug. After I made it read two auto‑generated markdown docs before each task and update them afterward, the AI started editing the right files, using correct tables, and never mis‑named anything. The whole workflow took a couple of hours to set up, and in just two weeks it’s saved me countless minutes and eliminated dangerous mistakes.

I tried to get Claude to handle my marketing—specifically to post automatically on social media—but it just wouldn’t do it. I kept prompting it to schedule and publish content, yet it stalled every time, leaving me to do everything manually. The experience was disappointing and felt like a waste of time, especially after hearing about its supposed all‑purpose capabilities.

I tried to wow my music‑teacher grandma by feeding the LLM photos of textbook exercises that a 12‑year‑old could solve in minutes. No matter how I re‑prompted or added context, it kept spitting out wrong answers. I felt embarrassed and reminded that true AGI is still far off, but at least my grandma now gets why students hand in incoherent work.

I love Claude’s web UI, but after paying for Pro and trying the desktop app with cowork and code features, it was a let‑down. Over three days I couldn’t get it to actually edit files or run workflows—most responses were “I don’t have access” or confidently wrong claims like “I updated the file” when nothing changed. It felt like extra work instead of help, leaving me frustrated and questioning what I’m missing.

I’ve been using Claude Code for six months to build a personal finance app—220k lines, Plaid integration, real users, and an upcoming App Store release. The AI helped me get the product together, but it also slipped serious security holes, rate‑limit bugs, and obscure edge‑case failures that proved costly to debug. The experience was a mix of amazement at what it could generate and frustration with the hidden flaws that only real‑world use exposed.

I was fed up with Alexa dropping Hindi commands and refusing to control my older gadgets, so I wired Claude in as the brain. I kept Alexa as the mic and speaker, but Claude reads device docs, parses intent, and runs the right script—WebSocket for the TV, DLNA for the set‑top box, RTSP→HLS for CCTV. Adding a new device is just a markdown file and a shell script. Commands like “Show CCTV on dad’s TV” or “TV band karo” now work flawlessly, turning a frustrating setup into a smooth, multilingual home hub.

I tried to give Claude a CLAUDE.md file with instructions, but it just kept ignoring it and kept prompting me to hit enter on everything. I was watching this time, so I noticed how it missed the guidance entirely, which was pretty annoying and made the whole process feel like a wasted effort.

I built a sticker app where users upload photos and Claude acts as the creative director. It analyzes each picture, identifies unique traits, and drafts nine tailored sticker concepts with detailed prompts for the image model. The results capture the dog’s real look, far better than generic prompts, though occasional over‑creativity needed guardrails. This workflow has been impressively useful for me.

I set up a workflow where I bounce ideas and bug fixes off Claude on my phone while I'm out. I talk through the problem, then ask Claude to log a detailed Linear issue. Back at my desk a daemon spots those “ready” tags, runs a Claude Code agent that writes, tests, and pushes the change. The simple tasks have been reliably auto‑implemented, letting me return to a ready‑to‑stage fix and feeling the tool’s assistance was surprisingly smooth and useful.

I tried using Claude Max 5x to tackle two long‑standing chores: reclaiming €600 from a plumber saga and resurrecting a beloved fantasy‑football game. The AI drafted a firm demand letter that got my money back in a week, and then built a full‑stack web app for the game in minutes. The experience felt astonishingly smooth, turning a frustrating legal grind and a daunting dev project into quick wins.

I set up a Godot MCP server and let Claude generate an entire tower‑defense game from a design doc. It created four tower types, upgrades, 20 waves, placement logic, UI, menus, and a game‑over screen—all using 36 custom tools that let the agent see and interact with the 3D scene. The click_at_world feature let it test placements just like a player, clicking world coordinates, taking screenshots, and fixing issues on the fly. I was genuinely impressed by how smoothly it turned my specs into a working game.

I spent two sessions with Claude Code and walked away with a fully‑functional tactical OSINT terminal. I described each feature, watched it stitch together feeds, then pointed out bugs like wrong legend colors and mis‑mapped icons—Claude corrected everything quickly. The result feels like a military‑grade map, and the brief‑generator even writes SITREPs locally. Building it felt surprisingly smooth for a non‑programmer.

I tried a tiny 100‑word prompt that asked Claude to tag my sentences as Intent, Reality, or Gap. Within minutes it reflected back that I’d been stuck in “Gap” for several messages, pointing out I hadn’t grounded any thread in reality. That mirror‑like feedback was oddly satisfying and instantly clarified why my project felt blocked, making the tool feel surprisingly useful.

I spent weeks wrestling with NAT traversal and Go’s networking quirks, and Claude turned out to be an absolute lifesaver. It helped me write the UDP multiplexing logic, debug edge‑case bugs, and get the protocol working end‑to‑end. The whole experience felt rewarding – the tool’s insight cut my frustration in half and let me finally ship a peer‑to‑peer stack that lets thousands of agents talk securely without any REST glue.

I used Claude’s code abilities to create Lockpaw, a free open‑source app that lets me lock my Mac screen while keeping AI agents running. The tool’s hotkey darkens the display, blocks input, and resumes with Touch ID—like a “Dog Mode” for Macs. Building it felt empowering; Claude’s help turned my design background into a ship‑ready product, and the project even gained unexpected GitHub stars.

I was frustrated by Claude’s confident but unfounded answers until I accidentally found Anthropic’s “Reduce Hallucinations” guide. Adding three simple system prompts—allowing “I don’t know,” demanding citations, and forcing direct quotes—made the output suddenly honest and source‑backed. I built a toggle to switch the strict mode on for research, and the difference felt like the tool finally respected the limits of its knowledge.

I keep getting the “Claude wants to use Updating file” prompt even after I’ve granted full folder access in the new Cowork desktop app. Every time I have to click “Always Allow” again, which interrupts my workflow. I’ve searched the settings but can’t find a way to make the permission stick, so I’m stuck repeatedly confirming just to keep working.

I spent weeks wrestling with Claude’s context – the session would drift, auto‑compact would wipe crucial details, and old dead‑ends kept polluting the conversation. Every new session felt like re‑priming for fifteen minutes, which was exhausting. Building Membase finally gave me a stable external memory graph, instantly restoring the right context and cutting the noise, so the tool finally behaved like a helpful partner.

I was thrilled to finally have a 1 M token context window for a massive bilingual audit project, but after the Claude desktop app update everything collapsed. The context shrank back to the default size, and my long‑running threads were forced into summarization, erasing crucial details I couldn’t recover. I’ve been stuck waiting for Anthropic support, feeling helpless as weeks of work vanished.

I used Claude Code to build BlushDrop, an anonymous love‑proposal app, even though I’d never touched Next.js or Supabase. The tool walked me through every step—designing a token system, writing rate‑limiting middleware, stripping EXIF data, setting up adapters, generating 97 unit tests, debugging realtime tracking, and handling Cloudflare/Vercel deployment. It felt like having an entire dev team; the result was a zero‑error, cheap‑to‑run product that works flawlessly.

I’ve been using Claude on the web to generate complex MySQL for my SaaS, feeding it dozens of SQL files and getting flawless results. When I try to replicate that via the API, I hit the context limit, have to trim everything, and the output drops dramatically. It feels weak and inconsistent compared to the web chat, and I’m stuck looking for a workaround.

I keep getting permission prompts every time the built‑in Explorer agent fires up, which feels like a nuisance when I just want to step away. I’ve noticed that sticking to a single session topic makes Claude act smarter, so I wish sessions had clear names like on the web portal. The Telegram integration was a disappointment – it hijacked all chats and seemed less organized, so I quit using it. I’m looking for tips from other CC users.

I tried to organize my thoughts using Claude as an AI canvas, turning each chat into a node that maps connections. The experience was eye‑opening but frustrating—whenever my prompts were vague, the output got messy, forcing me to write detailed specs and edge cases before any code worked. Once I tightened my instructions, the tool behaved much cleaner, but the need for meticulous prompting felt like a hurdle.

I tried Claude’s code verification tool by simply feeding it an image and letting it run without any back‑and‑forth. The experience felt surprisingly smooth – it parsed the visual data, checked the code, and returned clean results without me having to guide it through a maze of prompts. I was impressed by how effortless it was, and it saved me the hassle of manually reviewing every line.

opus 4.6

Terrible