AI Skills as loader spec, not prompts – why the architecture changes everything
Skills in AI systems are programs with structured execution stages, not static prompts, and their efficiency depends on how content is loaded across different runtime levels. The architecture of a skill—how instructions are distributed across metadata, procedural text, and on-demand references—determines performance and context usage more than the content itself. Understanding this loader-based design helps avoid common pitfalls like context bloat and model-specific failures.
- ▪A skill is a program with three execution stages: metadata loaded every turn, procedural instructions loaded on invocation, and references or scripts loaded on demand.
- ▪Treating skills as prompts leads to inefficiencies such as excessive context window usage and non-reproducible failures across environments or model versions.
- ▪Anthropic published the SKILL.md format as an open standard in December 2025, enabling cross-platform compatibility with tools like Claude Code, Kiro, Cursor, and Codex CLI.
- ▪Progressive disclosure in skill loading minimizes context costs by loading only necessary content at each stage, improving efficiency and scalability.
- ▪Level 1 (metadata) is critical for routing decisions, Level 2 (SKILL.md body) should be kept under 500 lines, and Level 3 (references and scripts) allows for scalable, on-demand content inclusion.
Opening excerpt (first ~120 words) tap to expand
What you're actually writing when you write a SKILL.mdINTERNALS.md #2 · Skills are programs, not prompts. How the skills runtime actually loads, and why the architecture is everything. Lax MeiyappanApr 30, 202611ShareA skill is a small program. It has three execution stages: 1\ what loads every turn, 2\ what loads on invocation, and 3\ what loads on demand. Because a skill is a program, it suffers from typical software rot—environment drift, version sensitivity, and silent, non-reproducible failures.You’ll see these failures in specific shapes. A skill that cost 20% of your context window, silently, before the agent did any work. A skill that worked perfectly until you shared it with a teammate, and ran the build in the wrong directory.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at Laxmena.
Discussion
0 commentsMore from Laxmena
