Token Secretary: Save Tokens
Paste the following prompt into your AI chat to install this skill:
Please follow https://skillhub.cn/install/skillhub.md to install @user_5b2fc4e3/token-sheng into your AI assistant.
About this skill
What It Solves
Token Secretary targets a concrete problem: cloud models often ingest long local knowledge context at once, increasing token cost and drifting from the source material. It fixes a retrieval → compression → cloud refinement pipeline so DeepSeek receives a filtered, compressed evidence block instead of raw documents.
How It Works
The skill reuses GLOBAL_RULES.md and existing local scripts:
- Intent classification: casual chat is answered directly; technical, process, and quote cases continue.
- Entity extraction: pulls materials, processes, parameters, and question type from the prompt.
- Dual retrieval: keyword search and FAISS semantic search run in parallel, then merge and keep up to 20 items.
- Local compression: llama3.2:1b compresses to under 500 characters; on timeout it falls back to the first 3 raw snippets.
- Cloud refinement: DeepSeek answers from the compressed block, with an observability marker at the start.
It depends on knowledge_querier.py, rag_semantic_search.py, compress_knowledge.sh, and the GLOBAL_RULES.md patterns. If FAISS, Ollama, or compression is unavailable, it degrades to keyword search, raw truncation, or pure cloud inference. It suits engineers with a local knowledge base who need controlled cloud token use, not fresh research without local sources.
Use Cases
- When answering material, process, and parameter questions, retrieve local keyword and semantic evidence, then compress it before cloud drafting.
- Before preparing a quote, extract material, process, and parameter clues from the local knowledge base into a short evidence block for DeepSeek.
- While maintaining an Agentic RAG setup, reuse GLOBAL_RULES intent, extraction, and fallback behavior instead of rewriting scripts.
- Before generating STEP suggestions, search local process notes and compress them to provide material context to the cloud model.
Best For
- Engineers with a local process knowledge base who need material, process, and parameter lookups
- Technical operations staff who compress retrieval results before asking a cloud model to draft quotes
- Automation engineers maintaining Agentic RAG scripts who want to reuse GLOBAL_RULES
- Process support staff who need material answers with lower hallucination risk
Related Skills
An AI workflow that structures personal knowledge bases, IMA, Yuque, and Feishu content into approved personas for paid WeChat mini-agent deployment.
Local workflow memory with matching and SOP updates.
An OpenClaw live streaming executor that initializes TRTC streaming, starts a real-time dashboard, generates viewer URLs, and continuously reports live events.
Breaks down physical supply chains for super-trends to identify second- and third-layer bottlenecks, runs valuation and reverse checks, and maintains trackable reports.