AWS has written the most practical guide yet to the part of agent skills nobody tests. In Best practices for writing AWS DevOps Agent Skills, Nisha Notani and Damini Kumar start from the 2 AM page: investigation quality “varies depending on who gets paged,” and skills encode the senior engineer’s workflow so the agent loads it automatically during an incident. The headline rule is about activation. “The description field in your frontmatter determines whether the agent loads your skill. A vague description means the agent skips the skill entirely, even if the instructions inside are excellent.”
The evidence is one anecdote and one customer, both labelled as such, and the anecdote is the useful one. A skill named db-health described as “Monitors database health” did not load when a production RDS instance hit max connections at 3 AM, because nothing in the description matched the alarm that fired. Rewritten as “Investigation procedures for RDS connection exhaustion, max_connections breaches, and DatabaseConnections alarm spikes,” with “same skill, same instructions inside,” it loaded every time. The second rule separates skills from agent instructions: AGENTS.md is “always on,” injected into every session with a hard 25 KB limit and about 120 lines recommended, while skills “load on demand.” Standing policy, such as “never surfacing secrets in plaintext,” goes in instructions; a procedure for “an ECS crash loop” goes in a skill. The third is that a skill without decision logic, “a checklist of nouns,” does not work.
The catalog reads AWS in two places, and this post sharpens a finding from yesterday. On the Amazon Web Services provider page, with 1,685 API pages and an agentic access profile of 5,181 operations, 3,518 acting, the agent skills dimension is unlit, alongside the MCP server. The skills AWS describes are written by customers into their own Agent Spaces, and the open-source DevOps Agent Tools repository the post links ships examples, but nothing on the record points an index at either. Amazon SES, which shipped skills for the AWS MCP Server last month, lights both.
The Kin Score is 59.4, strong band, with access clarity at 100.0 and contract quality at 26.3, and Agent Readiness is 30.5. The post’s best line applies to catalogs as well as agents. A skill is only useful if the thing looking for it can match its description to the problem in front of it. AWS has written that rule for the agent loading a playbook at 3 AM. The same rule decides whether an index looking for AWS’s skills finds any. On this record, the description that would let it match has not been written yet.