Agent Token-Saving Workflow Optimizer

Problem

In multi-turn, multi-tool, multi-document agent tasks, token waste often comes from repeated retrieval, verbose planning, dumping long tool output into context, and over-expanded step reasoning. This skill adds workflow constraints instead of replacing model capability: it makes output, loading, and validation decisions explicit to avoid useless tokens.

How It Works

• Routing: one-step tasks answer directly; two-to-three-step tasks plan internally then execute; tasks with three or more steps use plan -> pilot -> execute.
• Parallelism and deduplication: independent tool calls are issued in parallel; already retrieved data is not searched or read again.
• CoD reasoning: in compressible scenarios, each step stays within <=5 words, with the answer after ####.
• Context management: tool output above 500 tokens keeps key conclusions; long content may be written to workspace after user confirmation and referenced by path; complex tasks maintain progress.md.
• Memory injection: if missing memory is detected, it reads memory-inject.md, asks for confirmation, and writes to soul_md or agent_md.

Boundaries

Do not use it for simple Q&A or very short contexts. When anti-hallucination, high precision, creativity, first execution, or visible reasoning is required, validation takes priority over saving tokens. It does not modify conversation history; it mainly affects the agent output side. For data volume, aggregate no more than 200K characters and paginate 10-20 items per page.