Principle Guard The Context Window

Apply when context is filling up from large outputs, long files, or repeated reads. Filter bulk evidence and isolate bounded exploration when delegation earns its overhead.

HungryKelvin123 Updated

File contents

Guard the Context Window

The context window is finite and non-renewable within a session. Every token that enters should earn its place.

Why: Context overflow degrades reasoning quality, creates compression artifacts, and halts progress. Unlike compute or time, context spent inside a session cannot be reclaimed.

Pattern:

  • Filter before delegating. Narrow verbose outputs and reads locally first. For separable exploration that still needs substantial context, read ../david-mode/references/agent-runtime.md before dispatch. Require concise findings with exact evidence paths; the parent checks decisive evidence itself. If delegation does not pass its gates, continue sequentially.
  • Don't read what you won't use. Read selectively based on relevance. If a file isn't needed for the current task, skip it.
  • Keep frequently used content inline. Templates and references used on every invocation belong in the skill file, not in separate files that cost a read each time.
  • Size phases and cap scope. Limit files per phase, set turn budgets, account for mechanism costs.

HungryKelvin123/dstack/tree/main/plugins/dstack/skills/principle-guard-the-context-window commit d26dd8da4e

Frequently asked questions

npx skillmds@latest add hungrykelvin123/principle-guard-the-context-window