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 scope with per-phase file limits and turn budgets.

Use subagents whenever a payload is large and only its conclusions matter, such as verbose command output, screenshots, or long documents. The main thread keeps the summary while the subagent absorbs the raw data cost.

Every token entering the context window is non-renewable within a session. Repeated reads of the same content consume budget without adding information, causing compression artifacts and degraded reasoning later in the session.

No, if a template or reference is used on every invocation, keep it inline in the skill file. Separate files cost a read each time, while inline content is loaded once with the skill.

Context spent inside a session cannot be reclaimed, unlike compute or time. Summarization helps but introduces compression loss, so prevention through selective reading and payload isolation is more effective than recovery.