How do you save quota and tokens in Claude Code?

Put the Vercel team's 8 rules into AGENTS.md in the project root directory, requiring the simplest implementation, no premature abstraction, and preferring mature libraries, so the agent doesn't reinvent the wheel.

When tool output is too long, follow Pi's approach: keep only a short snippet in context, save the full result to disk, keep only the file path and offset position, and read it when needed.

Cutting out the middle content directly is risky, because it can't be recovered after trimming. For example, in a 30K-character test log where the error happens to be at character 12,000, it would be lost once trimmed.

When the context is nearly full, the system pauses for about 10 seconds to write a summary. One developer open-sourced a plugin that uses scoring to clear useless terminal logs while keeping core code and state.

For judgment tasks like selection, classification, and scoring, some people use dedicated models, saying they are cheaper and faster than having Codex or Claude act as judges, and that they make concurrency easier.

One team hosted Claude Code so that after hitting a limit it automatically waits for the reset before continuing, claiming they used nearly $10K worth of quota from a $200 subscription.