A lot of teams now know they should publish llms.txt.
That is useful. It tells the agent where the pages are. It does not tell the agent what job the product can finish, what inputs it needs, or which limits matter before it starts acting.
That second layer is where a capability page starts to matter.
A map and an operating note do different jobs
That is the clean distinction behind root skill.md for product capability discovery. A machine index is good for finding pages. A capability file is good for telling the agent what the product actually does and how to approach it.
It pairs naturally with publish llms.txt for agent retrieval. One file is the site map. The other is the field note. If you only ship the map, the agent still has to guess its way into the workflow.
Discovery should survive the real production path
The boring but important move is reverse-proxy forwarding for agent skill paths. A lot of teams generate the right files upstream and then lose them on the custom domain where the agent actually lands.
This is the same family of problem as agent-skills index for multi-workflow products. The value is not just writing the skill. The value is making the routes easy to discover from the hostname the buyer already trusts.
Trust gets better when the file is verifiable
I like agent-skills manifest with sha256 integrity for that reason. It is a small infrastructure detail, but it turns the capability layer into something closer to a real interface. The client can tell whether it fetched the expected instructions instead of a stale or mangled copy.
That matters more once the agent is about to use a live workflow. The same logic is behind skill frontmatter with compatibility and tool constraints. Clear constraints make bad automation less likely.
Access policy still decides whether any of this gets seen
The other quiet failure is access policy. Explicit AI-bot allowlist in robots.txt looks almost too small to mention, but it decides whether the discovery files get fetched and cited in the first place.
This is strongest for AI products, developer tools, SaaS docs, and creator software with a self-serve surface. If the product expects an assistant to help a buyer, a user, or an internal team, then the assistant needs more than a list of URLs. It needs the plain operating note too.
The activation move after that is simple: pair the capability page with one-command skill install from docs URL so discovery can turn into use without another layer of explanation.