Most people use AI to write code. I use Claude Code to run my life.
I've been using Claude Code as my primary terminal for a few months now on my Arch Linux desktop. Not as a coding assistant, but as an operating system for getting things done. Here's what that actually looks like.
Making memes from the command line
I type /meme and describe what I want. Something like:
"Nobody:" meme. The video should be a guy vibe coding on a bunch of monitors through the night. Find it on Instagram.
Claude Code searches Instagram and YouTube with yt-dlp, downloads the clip, probes the dimensions with ffprobe, builds an ffmpeg filter chain that stacks a white text bar on top of the video at exact 9:16 Instagram reel dimensions, and outputs a polished .mp4. The ffmpeg commands get pretty wild. It's calculating bar heights from text line count, forcing even pixel dimensions for libx264, matching frame rates between generated color sources and real video.
When I asked for 15 memes about my church staff and inside jokes, Claude ran five parallel yt-dlp searches, built memes in batches of three, and sent each batch to my iPhone over Tailscale as soon as it finished rendering:
for f in rr_rowdy6.mp4 rr_volunteers.mp4 rr_parking.mp4; do
tailscale file cp ~/$f iphone171:
done30+ videos produced and delivered to my phone in one session. No After Effects. No Canva. Just a terminal.
After that session hit 75% context, I told Claude to save everything it learned into a reusable /redrockscontent skill. All the ffmpeg templates, content formats, brand protocols, and church context, saved so next time I don't have to explain anything.
Filing a maintenance request through browser automation
My apartment complex has a clunky web portal for maintenance requests. Multiple dropdowns, cross-domain login redirects, cookie consent banners. I walked Claude Code through it once using the chrome MCP integration, which gives it control of a real Chrome browser.
Claude navigated to the portal, found the resident login, handled the redirect to the third-party auth provider, dismissed the cookie banner, and filled out the multi-step form. It selected "General" > "Pest control" > "Req indoor pest control trtmnt" from nested dropdowns. About 50 tool calls just for the browser interaction.
The funny thing: when it hit the password field, Claude just stopped and said "I need you to enter your password." After I logged in, it picked right back up like nothing happened.
Then I said: "Make this a skill so I never have to do this again."
Now I have a /maintenance slash command. I type /maintenance and say "my AC is broken" and Claude handles the rest. All the unit info, category mappings, and portal navigation are baked into the skill file.
Transcribing 82 voice memos across two machines
I had 82 iPhone Voice Memos going back to 2018. 1.1GB of audio sitting in iCloud. I wanted them transcribed for a personal project.
Claude SSH'd into my MacBook over Tailscale, copied all 82 .m4a files to my Linux desktop, installed Whisper, and started batch transcription with xargs -P4. When my GTX 970 proved too slow, it pivoted. SSH'd back into the MacBook, installed Whisper there too, and ran a second parallel transcription job on the M1 Pro.
Two machines transcribing at the same time, monitored from a single terminal. When a 5-hour and an 8-hour recording were blocking the pipeline, Claude killed them and re-ran the 24 missing files. End result: 69 memos transcribed, summarized, and compiled into a printable PDF workbook using Python's reportlab.
The SSH setup was pretty good too. When ssh-copy-id failed and I said the manual instructions were "too confusing," Claude spun up a temporary HTTP server on my Linux box and had me run one curl command on my Mac:
curl http://100.102.139.5:9999/id_ed25519.pub >> ~/.ssh/authorized_keys
Done. SSH worked immediately.
Building a personal assistant system in Obsidian
I wanted something like a personal operating system. Tasks, ideas, job leads, saved content, all stuff that Claude Code could read and write directly. After researching options (OpenClaw, Claude Cowork, iCloud sync daemons), the architecture ended up being the simplest possible thing: an Obsidian vault of plain markdown files.
Claude scaffolded the entire vault. 13 interconnected files with tables for Fiverr gig ideas, job leads, YouTube saves, contacts, and daily notes. Then it built a /pa skill that only loads on demand (I pushed back on putting everything in CLAUDE.md because it was too much context for every conversation).
Now I type things like:
/pa add fiverr gig: video editing for YouTubers
/pa save youtube https://... - great tutorial on AI agents
/pa what's on my list?
Claude reads the vault, appends to the right file, and confirms in one line. The vault is human-readable in Obsidian and machine-writable by Claude. Both interfaces work.
Running SEO audits and fixing my own site
My portfolio site had only 2 pages indexed in Google. I ran /seo audit https://brettridenour.com and Claude spun up parallel subagents to audit crawlability, schema, content quality, and sitemap structure all at the same time.
The root cause: a www vs non-www domain mismatch. Vercel served from www.brettridenour.com but every canonical URL, OG tag, and JSON-LD reference pointed to the bare domain. Claude fixed all 7 occurrences across 3 files, added AI crawler directives for GPTBot/ClaudeBot/PerplexityBot, created an llms.txt file, built the site, pushed to main, and verified every URL was correct on the live site with curl.
When I asked Claude to submit the sitemap to Google Search Console, it actually tried. Launched Playwright, navigated to GSC, but Google's anti-bot detection blocked the login. It just said "I can't get past this" and gave me manual instructions instead. Points for trying.
The pattern
None of this is really about code generation. Claude Code is a general-purpose agent that happens to run in a terminal. It chains together whatever tools the job needs (ffmpeg, yt-dlp, SSH, browser automation, web search, Python scripts, git) and when a workflow works, I turn it into a reusable skill.
Each skill is just a markdown file. No framework, no config, no build step. The /meme skill is 100 lines of ffmpeg templates. The /maintenance skill is 73 lines of portal navigation context. The /pa skill is a vault map with append-only rules.
The most useful software I run isn't an app. It's a prompt.