Skills Are the New Plugins: Procedural Memory Goes Open
For years, the way to make an AI agent do your job was the same way to make it do anything: you wrote a better prompt. Long ones, with step-by-step sections and "important: do not forget to…". It worked, but it was art, not engineering. The text lived in a chat box or a file nobody versioned, and the next agent couldn't reuse it.
That is changing. Around the middle of this decade the industry quietly converged on a file format for expertise itself: the Agent Skill. A skill is just a folder containing a SKILL.md file - YAML frontmatter holding a name and description, followed by plain-markdown instructions - plus optional scripts/, references/, and assets/ subfolders. That is the entire spec, and that is the point. It turns "how I do a thing" into something you can commit to a repo, review in a PR, and hand to any agent that speaks the format.
The format that went cross-client
The standard, published at agentskills.io, was originally developed and released by Anthropic, then opened to the wider ecosystem. Because a skill is just a directory, adoption spread fast: Claude Code, Cursor, Codex, ChatGPT, and a growing list of independent agents all support it. The discovery paths are conventionalized too - a project can drop skills in .agents/skills/, and a user can keep their own in ~/.agents/skills/ so a skill follows you between tools.
This is what "skills are the new plugins" means. Plugins wired a tool into one specific product; a skill carries the knowledge to use many of them, and is portable by design.
What makes Hermes different: the learning loop
Most agents can load a skill you give them. Hermes Agent, the open-source agent from Nous Research, went a step further: it writes its own. Its core pitch is a closed learning loop - after finishing a complex task, Hermes can autonomously distill what it just did into a reusable skill, and it refines its existing skills during use. A skill stops being a one-way import and becomes something the agent maintains, like a senior engineer updating an internal wiki because they'll hit this again.
On top of that, Hermes came in early on standardizing the ecosystem. Its skills are compatible with the agentskills.io spec, and its documentation positions skills as "procedural memory" - loaded only when a task calls for them, not dumped into every prompt. That progressive-disclosure model keeps the context window lean even when you carry hundreds of skills around. The bundled Skills Hub points at over 88,000 skills across every major registry, so a fresh install doesn't start from zero.
Why it matters
Two shifts here are bigger than any single tool. First, expertise has become a buildable artifact. The hard-won steps for reviewing a PR, running a migration, or formatting your company's spec sheets now live in files with diffs, owners, and tests. Second, your accumulated knowledge stops being locked in one vendor. Write a skill to the standard, and it survives whatever hyped agent you're using next.
The irony is that the "just write a better prompt" era built an entire cottage industry on prompt engineers selling secrets. The open standard turned that secret sauce into something boring and wonderful: a plain folder in your repo. That is how real infrastructure tends to happen.