context-budget

context-budget audits how an agent is spending its scarcest resource — the context window — and trims the overhead. Less noise in the window is not just cheaper; it measurably improves the quality of what the agent produces.

What it solves

A degrading agent is usually a full agent. As the window fills with stale output, re-read files, and verbose configuration, the signal-to-noise ratio drops and responses get vaguer, slower, and more likely to lose the thread. The failure looks like the model getting worse; it’s actually the context getting worse. This skill finds where the tokens are going and reclaims them.

How it works

It treats the context window as a budget to audit, not an infinite buffer:

• Measure — identify what is actually consuming the window: oversized configuration, repeated file reads, verbose tool output, instructions that could be a rule instead.
• Trim — cut the overhead that doesn’t earn its place, so the working set is the relevant material and not the accumulated exhaust.

The payoff is twofold: lower token cost, and higher answer quality — a focused window keeps the model on the actual task.

When to use

When responses start degrading (the tell-tale sign of context pressure), or proactively when setting up a project so the per-session overhead is low from the start. It sits in the context-* family next to context-handoff.