principle-guard-the-context-window

Route verbose outputs, screenshots, and large documents to subagents so the main thread receives only summaries. Read files selectively based on relevance, and cap the number of files per phase with explicit turn budgets.

Use subagents whenever a payload is large and only its conclusions matter, such as log dumps, big documents, or fan-out searches. The subagent processes the raw data and returns a condensed summary to the main context.

Content used on every invocation belongs inline in the skill file, because separate files cost a read each time. Only infrequently used, in-depth material should live in external reference files.

Context overflow creates compression artifacts and forces truncation, so the model loses access to earlier details. Unlike compute or time, context spent within a session cannot be reclaimed, so every token must earn its place.

Summarization loses detail, so decisions needing raw data must happen in the subagent that read it. These techniques reduce consumption but cannot expand the fixed window, so extremely large tasks still require careful phase sizing.