jcorpac Meta-Context Budgeting
AI models have a "budget" (the context window). Every token you use leaves less room for reasoning.
Budgeting Strategies
- Purging: Identifying and removing outdated logs, repetitive instructions, or completed task summaries once they are no longer needed.
- Summarization: Compressing long technical discussions into concise action points before proceeding to the next phase.
- Priority Filtering: Keeping "System Instructions" and "Foundational Constraints" at the highest priority, while ephemeral chat history is lower priority.
Tools
- Token Counters: Using tools to monitor current usage.
- Selective Retrieval: Only fetching the most relevant "chunks" of a codebase or skill library.
Best Practices
- Clean State: Occasionally starting a fresh session and only carrying over the "Essential State" (ADRs, current plan, critical constraints).