I’ve been publishing my Claude Code skills to a marketplace I set up. I call them powerups, inspired by Jesse Vincent’s Superpowers.

Some are good. Some are held together with duct tape and optimism, and I can tell which is which by now. Some are in use by others, most are just for my own use. I’m making the repos public so you can see what I work on, should you care.

The latest is rate-my-cli. You point it at a CLI’s source and it scores how well that tool serves an agent — 47 pass/fail checks across ten principles. I stole the principles from Trevin Chow’s essay on agent-native CLIs and then kept adding my own every time a real project embarrassed me. It reads the code (never runs it — that’s how you avoid a stray delete eating your data), hands back a scorecard with file:line evidence and a remediation plan sorted by how much each gap hurts an agent, and it’ll fix the small stuff for you. The subsystem-sized gaps it just points at and lets you decide.

We spent fifty years making command-line tools friendly to humans. Now half my tool use is an agent shelling out to them, and it turns out “friendly to humans” and “friendly to agents” aren’t the same thing — output you can’t parse without regex and a prayer, errors that don’t say what went wrong, exit codes that are always zero. rate-my-cli is me trying to see my own tools the way the agent sees them, which is usually less flatteringly than I’d like.

In the distro I include some analyses of CLI tools I use. I vibed something for Azure DevOps (azdo), something for Clickup (cu), and something for Microsoft 365 (go365). Yes, I am trapped in Windows corporate land. Send a St Bernard with a Raspberry Pi around its neck. They all gained from being reviewed through this lens.