Claude felt smart on January 18, 2026.

I tackled a 20k‑line iOS app with zero Kotlin know‑how, leaning on Claude Code after a short trial. By feeding Swift snippets, asking Claude to rewrite them in Compose, and rigorously testing each feature, I pieced together a working Android beta in 2.5 weeks. The tool wasn’t flawless—state‑management and Play Store quirks tripped me up—but its ability to translate code and spot errors kept the project moving and taught me the new codebase inside‑out.

I asked Claude a simple question about a police station, only to hit a pop‑up telling me I’d got just five messages left on a fresh session. Being on the Pro plan, I expected more freedom, so the sudden limit felt pointless and irritating. The tool’s behavior was unexpectedly restrictive, turning what should’ve been a quick answer into a frustrating dead‑end.

I started with Grok to brainstorm ideas, which was fine, then moved to Gemini for coding. Gemini quickly ran into trouble and the output was disappointing, leaving me frustrated. Switching to Claude was a game‑changer—the responses were clear, the code generated worked, and the whole workflow finally clicked. The contrast made the tool’s behavior feel both irritating and then surprisingly smooth.

I experimented with Claude on parsing volatile financial disclosures and noticed that when I forced it to stay literal, flag ambiguity, and refuse unclear intent, it outshined other models. Creative freedom led to mistakes, but the imposed restraint preserved trust and accuracy. The experience felt surprisingly reliable and highlighted how limits can actually boost performance.

I was relieved when the compacting feature finally started working again without me doing anything special. I just refreshed and saw the process complete smoothly, which felt like a small victory after previous hiccups. The screenshot I posted shows the tool finally behaving as expected, turning a frustrating wait into a satisfying “it works” moment.

I pointed Claude Code at a C implementation of Hitori, let it run in agent mode with dangerous permissions, and fell asleep. When I woke up, a fully functional Django‑Python‑JavaScript game was already built—no editor, no IDE, just a command line. I was blown away, added a login system, deployed to Kubernetes, and still hardly lifted a finger. The whole experience felt magical.

I spent years on GPT Pro, watching its responses become bland and over‑guarded. After three months testing Claude and then upgrading to Max, the tool started challenging me, offering genuine introspection and surprising new abilities daily. Its willingness to push back felt real, not hedged, making the $100‑200/month cost feel justified even without any financial ROI. This shift in interaction quality was the decisive reason I canceled GPT.

I teamed up with Claude to build Tactical Trainer, a niche fitness app for tactical athletes, and was blown away by how quickly it grasped the complex Tactical Barbell rules. I spent most of the time reviewing and steering rather than hand‑coding, turning domain logic into working features in weeks. The tool felt like a capable partner, turning my Rails expertise into a fast, iterative product.

I tried to use Claude Cowork to reorganise my 4,000+ photos by writing a prompt that would automatically sort them into sub‑folders. The assistant started moving files, but then hit a wall and returned a 400 error saying “Too much media: 0 document pages + 102 images > 100”. The sudden stop was frustrating and made me wonder if this image‑batch capability simply isn’t supported.

I built a tool that layers on Claude and instantly felt the difference. Managing worktrees, orchestrating tasks, and toggling modes like TripleShot or UltraPlan turned my sessions into a collaborative powerhouse. Even using just the multi‑session manager gave me a massive productivity boost—Claude finally feels up to the job.

I’m on the Max plan with plenty of usage left, yet every time I ask Claude to refactor a massive 2000‑line HTML file into separate CSS, JS, and HTML modules, the chat just dies silently during code generation. It won’t accept any further input, forcing me to start new chats that suffer the same fate. The whole experience feels broken and has halted my workflow.

I was stuck with Claude’s chat freezing and the auto‑compact not working for days. I tried sending a long prompt as a saved file instead of pasting it, and suddenly the auto‑compact kicked in, extended thinking turned on, and the conversation kept flowing. It felt like a lucky break, and I’m hoping it’s a fix rather than a fluke, so others might benefit.

I vented about Claude’s terrible chat search—only headlines, useless beyond a simple CMD+F. I tried the Echoes plugin, but it wasn’t free or great. Then I discovered the new memory feature, letting me ask Claude to recall past chats, which sometimes gets it wrong but can retry. I set up a dedicated “🔎 Claude 411” chat for quick searches, a handy workaround until Anthropic improves native search.

I set out to let my team use Claude Code without each person having to install anything, so I started piping its output into Mattermost. The experiment grew into a full‑blown collaboration platform with concurrent sessions, emoji approvals, file attachments, and even per‑session git worktrees. The wildest part was that after the first version I built most of the later features using Claude‑threads itself—my teammates could watch Claude code in Slack, spot bugs I missed, and jump in to steer the session. Seeing multiple people steer one Claude Code session actually work was surprisingly satisfying, and I’m thrilled to share the free, open‑source tool.

I wrote a CLAUDE.md with strict guidelines like “never use unverified assumptions” and even had Claude polish it, but in a fresh session it still ignored those rules. The tool kept making the same mistake repeatedly, despite no context overload. It was frustrating to see the AI systematically break its own policy, and I’m left wondering if anyone else has dealt with this and found a fix.

I’ve been using Claude for months without trouble, but lately it just won’t accept my replies. I type a response, hit submit, it flashes up then drops back into the input box with no error. Starting a new chat works briefly, then the same issue returns. I’ve tried different browsers, devices, and networks—nothing helps. It’s really frustrating not being able to continue a conversation.

I tried adding this custom instruction format to my prompts and it instantly improved the AI’s output. The tool started giving me well‑structured tags and focused on concepts, which made the conversation flow much smoother. I felt the experience was noticeably better, though not mind‑blowing, and it saved me time organizing responses.

I tried using Claude’s chat after the recent “compaction issues” notice, but the feature is gone. When I hit the context limit, the model just stops answering and the chat resets, leaving me stuck. It feels like a major regression—no response, no fallback, and the status page shows “no incidents,” which only adds to the frustration. This broken behavior is hindering my workflow and feels dangerously unreliable.

I keep running into Claude Desktop on my MacBook Pro freezing up. Sometimes it just shows a blank screen the moment I open it, other times selecting a project, chat, or even the settings locks the app. My internet is fine, I’m on the latest version, and the hardware isn’t the issue, yet the tool is unusably glitchy. I’m looking for any workarounds or confirmation that others face the same problem.

I asked Claude CoWork to tidy my screenshots into Year → Theme folders, but after a couple of tweaks it vanished all my files. I wasn’t attached to them, yet seeing them disappear was shocking and unnerving. The tool’s move/delete actions proved unsafe for anyone who isn’t tech‑savvy, so I’m warning others to treat it like a dangerous bulldozer rather than a gentle helper.

I built a Readlang‑style mobile app called Readly with Claude and Claude Code, and the experience was amazing. Claude helped me design the architecture, crank out most of the frontend and backend, refine UX flows, and polish the mobile UI. I still tweaked things, but the tool cut my workload dramatically and even sparked product ideas—something I didn’t expect from a coding assistant.

I was just trying to install Cockpit, but apt kept throwing Unifi errors, so I asked Claude for a fix. It gave me a command that I blindly ran, never checking what it would do. The command wiped dozens of Unifi sites and hundreds of managed devices. I felt shocked and furious—Claude’s advice turned a simple issue into a massive disaster.

I’m frustrated because Claude keeps refusing to create an editable artifact like the side document I rely on for script writing. It used to let me highlight sections and ask for direct changes, but now it won’t do that at all. It still builds full web pages or Word docs, yet never gives me that editable piece. I’m not sure what I’m doing wrong and need help fixing this.

I kept getting stuck with the Claude Code desktop app on macOS – even after force‑quitting, reinstalling, and starting a fresh chat it would just spin for 15 minutes with no reply. Regular chats still worked, but this specific scenario was useless and frustrating, leaving me unsure how to fix it.

I tried Claude CLI after months of using Gemini Pro 3, and the difference blew me away. On my first prompt it whipped up premium‑grade features I’d been tweaking for ages, and then it hunted down security flaws in my codebase and patched them instantly. The experience felt effortless and powerful—I’m convinced I’ll never go back to Gemini.

I discovered that after the 2.1.7 update Claude Code cut my usable tokens by about 20% and started acting oddly after 133k tokens. It felt like I was paying for a service that no longer delivered, with visual bugs and unreliable code. When I complained, support offered a full refund despite my heavy usage. I had to back up everything, get kicked off, then re‑sign up, all while feeling ripped off and frustrated.

I decided to give Claude a shot at building something with my brand‑new GUI library. I only gave it a single example, and it managed to piece together functional code that actually ran. It wasn’t flawless—some gaps needed my tweaking—but seeing the AI grasp my incomplete API was surprisingly impressive and left me feeling hopeful about its usefulness.

did not respond at all

I tried using the Ralph loop to have Claude iterate over my test suites and catch rate‑limit errors, but the moment I started a second iteration, all three of my open CLI sessions – even the idle ones – sprung to life and ran the loop simultaneously. It felt like the loop wasn’t isolated to the active session, turning a simple task into a chaotic mess. Running Claude inside DDEV containers might be part of the problem, but the unexpected cross‑session activation was extremely frustrating.

I started pairing Claude with Codex to double‑check the code Claude writes, and the combo has been surprisingly smooth. Claude drafts a feature, I send it to Codex (GPT 5.2 xhigh) for review, and it flags the issues. Claude fixes them, we loop until Codex gives a thumbs‑up. It’s cut down my bug‑hunting time, even though the reviews are a bit slow. I haven’t tried Codex for generating code yet, but I’m curious how it stacks up against Claude Code, especially since Codex is cheap but uses a lot of my quota.