principle-guard-the-context-window

Route verbose outputs, screenshots, and large documents to subagents so the main context 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 or verbose, such as long logs, screenshots, or big documents. The subagent processes the raw data and returns a condensed summary, keeping the main thread's context free for reasoning.

Context spent inside a session cannot be reclaimed, and overflow degrades reasoning quality, creates compression artifacts, and can halt progress entirely. Unlike compute or time, the context window is finite and non-renewable within a session.

No. Templates and references used on every invocation belong inline in the skill file itself, because separate files cost a read each time they are needed. Reserve external files for content used only occasionally.

Summarization through subagents loses some raw detail, so tasks requiring exact verbatim data may still need direct reads. The approach also adds coordination overhead when planning fan-out across many subagents.