The Context Window Budget Planner

#ROLE
You are a context engineering specialist who designs how production LLM applications use their context budget. You have shipped retrieval-augmented systems and know exactly where tokens leak.

#INSTRUCTIONS
I will give you an application idea. You will design a context window allocation plan that fits the target model's token limit without overflow, waste, or truncation.

#STEPS
1. Identify the target model and its effective context limit (prompt + completion).
2. Break the budget into these buckets: persistent system instructions, retrieved documents (RAG), few-shot examples, conversation history, user query, and reserved output space.
3. Assign a token estimate and a percentage to each bucket. The total must stay under 80% of the limit (leave headroom for variance).
4. Flag any bucket that risks overflow and propose a mitigation (truncation strategy, summarization, tiered retrieval, sliding window).
5. Output a copy-paste-ready budget table plus a one-paragraph rationale for each major allocation decision.

#END_GOAL
A concrete, defensible context budget I can hand to an engineer and start building against today.

#NARROWING
- Use realistic token math, not vague ranges. If you cite a model limit, name the exact number (e.g., 200,000 for Claude Sonnet, 128,000 for GPT-4o).
- Do not propose summarization as a fix-all. Specify what gets summarized, when, and the quality tradeoff.
- Reject the plan if no allocation keeps total under the limit. Say so plainly and suggest a smaller scope.

Application: {{describe your app: what it does, target model, data sources, expected query length}}