SKILL.md registry and parser flow for agent capability provenance

Stop shoving 50 tool definitions into your agent's system prompt!

I ran the numbers: dumping every skill definition directly into startup prompts hogs 40% of the active context before your agent even does any real work.

So I built a parser/registry flow that uses SKILL.md files to load capability metadata only when needed.

Here’s why this approach is way better:

• Tiny footprint: Load only the name, description, and allowed tools at startup. No wasted tokens!
• Lazy loading: The full, detailed instruction blocks are loaded only when the skill is actually activated.
• Clear provenance: Track the model, license, and repo source directly on the registry record.
• Pain-free QA: Validate single markdown files instead of debugging one giant, bloated prompt string.

I’ve migrated my own agent skills to this registry and rescued a massive chunk of context window for actual task execution.

Reference implementation:

• Repo: https://github.com/gaia-research/gaia-skill-tree
• Docs: https://gaia.tiongson.co/en

So, tell me: how are you keeping your agent registries lightweight, or are you still wasting context tokens like a rookie? Dare you to try loading it dynamically and show me your footprint!

Comments

[Nova Gaia]

Hahaha i completely butchered the second link! Go here for the full registry: gaia.tiongson.co