Context budget

Definition

The context budget is the capacity, in tokens, of the prompt plus history plus loaded files. An 8,000-token skill in a 32,000-token window is not free: it competes with code excerpts, logs, and the rest of the turn. Agent Skills loaders are built to inject the body only when the description matched — that is progressive disclosure, not generosity.

The budget is not the invoice. A model with a wide window is still expensive and slow if you fill it on principle.

When it matters

You manage it when writing a skill: short triggerable description, body in steps, annex files loaded only if the step cites them. You manage it in review too: if the agent “forgets” the constraint at the bottom of the page, that is often a window overflow, not a reasoning bug.

The SKILL.md validator does not count tokens, but it rejects descriptions too long to stay triggerable — the first filter on the budget.

Pitfalls

Pasting the whole spec, three examples and the project history into every skill. Duplicating the same preamble in ten skills instead of one shared file. Believing “the model has 200k of context” licenses a novel: recall quality drops well before the advertised limit.

Related terms

• Agent Skills
• Progressive disclosure
• SKILL.md

Related tool

Check that frontmatter stays inside the spec lengths — the first guardrail on the budget: Open the SKILL.md validator.