Agent Skills: Disseminating Expertise

Essay by Tristan Handy. Reflects on dbt agent skills as encoded expertise -- not documentation, not MCP tools, but procedural guidance that AI agents load dynamically.

Curated Topics

• Skills as encoded expertise: 12kb of markdown encapsulating hundreds of hours of collective migration experience; the migrate-to-fusion skill performed a full dbt Core 1.10 migration with zero human help
• Skills vs MCP tradeoff: MCP injects full tool schemas whether relevant or not (44k tokens); Skills + CLI approach is 10-32x cheaper and hit 100% task completion vs MCP's 72% in one benchmark
• Skills package manager landscape: Three leaders -- Vercel/skills.sh, Tessl, SkillsMP. Active discussion about dbt's package manager building native skills support (dbt deps installs packages + skills)
• Technical knowledge vs best practices knowledge: Documentation answers "how does this work?" but skills can also encode "how should you use this?" -- the patterns, traps, and judgment calls that experienced practitioners share informally
• Transformative vision: dbt packages bundling their own skills (e.g., dbt-datavault includes Data Vault best practices the agent auto-deploys). OSS expertise, not just OSS software
• Writing for agents vs humans: Agent-targeted content can be more declarative ("do this") vs human content requiring space for opinions and context

RDCO-Relevant Links

• Skills as consulting delivery mechanism -- we could package RDCO dbt best practices as agent skills
• dbt-datavault + skills bundling -- relevant to SDG pipeline architecture recommendations
• MCP vs skills economics -- informs our tooling recommendations for client agent setups