Skills are how you give AI coding agents new capabilities – folders of instructions, scripts, and references that agents discover and use. The format is defined by the Agent Skills spec (agentskills.io), but writing good skill metadata is surprisingly hard. Bad frontmatter leads to "skill collision" – the agent picks the wrong skill for a prompt.
Sensei is a linter and iterative improver for skill files. It runs programmatic checks against the spec (name format, field constraints, required fields) plus advisory checks we derived from empirical testing: trigger phrase detection, anti-trigger coverage, token budget tracking, routing clarity for MCP tools, and over-specificity warnings.
It works as a skill itself – you say "run sensei on my-skill" and it loops: read → score → improve → test → repeat until the skill reaches target compliance. Think of it like eslint but for agent skill metadata.
Skills are how you give AI coding agents new capabilities – folders of instructions, scripts, and references that agents discover and use. The format is defined by the Agent Skills spec (agentskills.io), but writing good skill metadata is surprisingly hard. Bad frontmatter leads to "skill collision" – the agent picks the wrong skill for a prompt.
Sensei is a linter and iterative improver for skill files. It runs programmatic checks against the spec (name format, field constraints, required fields) plus advisory checks we derived from empirical testing: trigger phrase detection, anti-trigger coverage, token budget tracking, routing clarity for MCP tools, and over-specificity warnings.
It works as a skill itself – you say "run sensei on my-skill" and it loops: read → score → improve → test → repeat until the skill reaches target compliance. Think of it like eslint but for agent skill metadata.