micsapp Openrouter Context Optimization
Overview
This skill covers techniques for managing token budgets, optimizing prompt length, and making the most of each model's context window.
Prerequisites
- OpenRouter integration
- Tokenizer library (tiktoken for OpenAI models, or approximate counting)
Instructions
- Count tokens before sending: Use
tiktoken(Python) orjs-tiktoken(JS) to estimate token counts before making API calls; abort if the prompt exceeds the model's context limit - Implement conversation pruning: For multi-turn chats, drop or summarize older messages when total tokens approach the context limit, keeping system prompts and recent messages
- Optimize system prompts: Move static instructions to the system message and keep them concise; every token in the system prompt is repeated on every request
- Use response length controls: Set
max_tokensto the minimum needed for your use case; shorter completions cost less and return faster - Select models by context need: Route short-context tasks to cheaper models and reserve large-context models (128K+) for tasks that genuinely need them
Output
- Token budget calculator that warns before exceeding context limits
- Conversation pruning middleware that keeps chats within context bounds
- Cost reduction from optimized prompt engineering (typically 20-40% savings)
Error Handling
| Error | Cause | Fix |
|---|---|---|
400 context_length_exceeded |
Prompt + max_tokens exceeds model limit | Prune messages or switch to a model with a larger context window |
| Truncated responses | max_tokens set too low |
Increase max_tokens or split the task into smaller requests |
| Token count mismatch | Wrong tokenizer for the model | Use the correct tokenizer per model family; fall back to rough estimation (1 token ~ 4 chars) |
See ${CLAUDE_SKILL_DIR}/references/errors.md for full error reference.
Examples
See ${CLAUDE_SKILL_DIR}/references/examples.md for runnable code samples.