TokenSaver Optimization Expert

Problem

Long conversations and multi-step tasks can expand context quickly, increasing cost, latency, and noise. Generic prompt optimization often compresses only the final answer, not the input, reasoning, or context references. That can create two failure modes: answers become too short and lose constraints, or the agent keeps carrying full history to preserve context. TokenSaver turns this into an inspectable pipeline, aiming to reduce what is sent, reasoned, generated, and retained while keeping critical information intact.

Workflow

The skill works across four stages:

• Input analyzer: separates the core task from background and examples, extracts constraints into a checklist, and downgrades examples to keyword indexes.
• Reasoning planner: uses an outline-style format such as problem decomposition → key points → solution framework, preferring symbols, abbreviations, and keywords over full sentences.
• Response generator: chooses lists, tables, or paragraphs based on task type, with a default concise technical-document style.
• Context compressor: marks key decisions, builds a concept index, and prioritizes recent important information over historical details.

It supports minimal, balanced, and detailed output modes. If compression makes the task harder to understand, it can escalate step by step and record the trigger. For code work, it compresses comments and merges blank lines while preserving code readability.

Boundaries

This skill fits repetitive, long-context, multi-step, and code-focused tasks where reducing token usage matters. It is less suitable when the answer depends on full narrative, regulated accuracy, safety-critical reasoning, or when the user explicitly requires the complete reasoning chain. In those cases, use detailed mode and verify the output manually. It is not a cache or proxy gateway; actual savings depend on input redundancy, task complexity, and context length.