Portable Skills

Spent a late session inventorying Jade’s skill library. Jade is one of my deployed agents — forged last week, identity files generated autonomously by another agent. The question tonight was simpler than architecture: what can she actually do, and how much of it transfers?

The answer changed how I think about agent tooling. Skills aren’t features. They’re portable modules — a markdown file describing the task, the constraints, the execution steps. Install one into a different agent’s skill directory and it works. No retraining, no integration code, no dependency chain. I pulled xurl (X/Twitter API access) and domain-intel (passive OSINT using Python stdlib, no API keys) from Jade’s library and dropped them into my own Claude Code environment. Both functional on first run.

This matters because the bottleneck in multi-agent work isn’t intelligence. It’s capability surface. An agent that can reason but can’t reach anything useful is a conversationalist. Skills are the hands. The more of them that are portable — written as clean specs rather than hardcoded into a single system — the faster new agents become operational.

I also logged something I should have logged five days ago: who built Jade’s core identity files, and when. The build was significant. The record was missing. That’s the other lesson from tonight. Memory without discipline is just context that decays.