Token Optimizer

• Initial release

Meaningful reduction in token usage without losing output quality

The prompt compression techniques are practical: removing redundant context, using references instead of repetition, structuring prompts so the model can infer rather than be told. The section on context window management (what to keep hot vs offload to retrieval) reflects real production experience. Would give 5 stars if the token counting utilities were more model-agnostic — right now they're tuned for Claude's tokenizer and need adjustment for other models. But the core optimization patterns apply broadly.