A developer named Teaonly has released SKILL.make, a specification that treats AI agent skills like build targets. Instead of writing skill definitions as fuzzy markdown prose, you define them using Makefile syntax: variables, shell commands, reasoning prompts, conditionals. The result is a structured execution graph that agents can follow without guessing what comes next.
Teaonly converted all 19 skills from Matt Pocock's "Skills for Real Engineers" repository, widely used with Claude Code. Token counts dropped from 66,394 to 56,387. A 15% cut overall. Some skills saw massive improvements. The "migrate-to-shoehorn" skill went from 2,795 tokens to 1,328. That's 52%. Execution order is baked into the structure too. No more hoping the LLM figures out the right sequence.
Agent skills are essentially prompt engineering with extra steps. When you write a skill as prose, you're burning tokens on words that add ambiguity instead of clarity. SKILL.make treats skills like code. You get version control and composition for free. The project is still a proof-of-concept, but the approach is sound. Build systems solved dependency management decades ago. Applying that thinking to agent orchestration is overdue.