I went looking for a plugin yesterday and found an entire economy I didn't know existed.
There are now multiple marketplaces for Claude Code skills and plugins. Not five or ten listings. Hundreds. One repo alone has 346 plugins with over 1,900 validated skills. There's a grading rubric. Community contributors. Categories like "DevOps Automation" and "Executive Assistant" and "Overnight Autonomous Developer."
That last one is a skill pack that runs your codebase for 6-8 hours while you sleep, writing tests the whole time so it doesn't break anything. Someone reportedly replaced a $4,000/month executive assistant with another plugin. I can't verify that claim, but the fact that it's even plausible tells you where things are headed.
Here's what caught my attention: these aren't code libraries. They're behavior files. Most of them are just markdown with instructions. No build step, no dependencies, no runtime. You drop a file in a folder and your AI starts acting differently.
ccpi install overnight-devThat's it. Now your terminal knows how to run an autonomous coding session with test-driven guardrails.
I spent a few hours evaluating which packs made sense for my setup. Installed some, rejected others. The interesting decision wasn't which tools to add -- it was the architecture question. Some systems want you to switch between "modes" -- a coding mode, a writing mode, a research mode. Each mode loads different context and skills.
I rejected all of them. I want one agent with everything loaded. Flat, not modal. Same way I don't switch between a "writing computer" and a "coding computer." It's just my computer.
That's a bet. More context means more noise. But it also means the agent can connect things that siloed modes never would -- pull SEO data into a blog post, reference a project timeline while answering a technical question, make a meme about something it helped me debug yesterday.
The ecosystem is moving fast. A month ago I was hand-writing every skill from scratch. Now there are curated registries, install CLIs, quality scores, and competing marketplaces -- the whole infrastructure of a software ecosystem, except the "software" is just instructions for how an AI should behave.
We're packaging cognition now. That's new.