principle-guard-the-context-window

Route verbose outputs, screenshots, and large documents to subagents so the main context receives summaries rather than raw payloads. Also read files selectively and cap the number of files processed per phase.

Use subagents whenever a task produces large outputs or requires reading many files whose full contents are not needed in the main thread. The subagent processes the bulk data and returns only a condensed summary.

Context overflow creates compression artifacts and crowds out relevant information, since context spent within a session cannot be reclaimed. Keeping the window lean preserves reasoning quality and avoids stalled progress.

Templates and references used on every invocation belong inline in the skill file. Separate files cost an extra read each time, consuming context that could be spent on the actual task.

Summarization loses detail, so subagent summaries may omit nuances needed later. These techniques manage consumption but cannot expand the fixed window, so extremely large tasks still require careful phase scoping.