On this page

Existence ≠ content

Knowing that an atom for "OWASP input-validation rules" exists is different from holding all 1,841 bytes of its prose in working memory.

Wikipedia separates the two. Every reader of the encyclopedia knows there is an article on prime numbers — but they don't have its 4,000 words in their head until they navigate to it. The Wikipedia index is small; the articles are large; readers pull the article on demand.

Pre-Skill-Wiki agent stacks routinely conflate the two. Every skill that might be relevant ships its full prose into the prompt every turn, "just in case." This is the design pivot Skill Wiki rejects.

How Skill Wiki separates them

The protocol distinguishes two surfaces that an agent interacts with:

Surface	Cost	Always in context?	Contains
Index (`_index.xml`)	~3 KB regardless of corpus size	Yes	id, kind, one-line summary, edges out
Atom body (chunks/*.md)	~30 / ~150 / ~380 tok per atom	No — loaded by id, on demand	full description, parameters, examples

The index is the agent's map; atom bodies are the territory.

What this enables

Bounded cost

Token cost grows with what the agent actually decides to load, not with corpus size. A 1,000-atom corpus costs the same to "have available" as a 50-atom one. The marginal atom is free until used.

Bounded pollution

Bad atoms in the corpus don't poison every turn. They sit in the index, get filtered out by retrieval per-turn, and never enter context unless retrieved. A bad SKILL.md, by contrast, pollutes every turn it's loaded for.

Push → pull inversion

The traditional pattern is push: the system pushes everything the brief might need into context, hoping the model picks the right parts. Skill Wiki is pull: the index says "an atom for GDPR right-to-erasure rules exists"; the agent decides whether to read its summary, core, or full projection.

Why "exact" is exact

The Wikipedia analogy isn't approximate; it's structural. Wikipedia's Special:AllPages is a 7 MB index. Loading it doesn't mean loading the encyclopedia. The MediaWiki runtime gives readers stable URLs, and articles materialise when the URL is requested. Replace "URL" with "atom id," "browser" with "agent," "MediaWiki" with "MCP server" and you have Skill Wiki.