industrial-ai-troubleshooting-agent: Instructions file for Claude Code

Boiler Historian — System Reference

This system connects an LLM to four data sources for a coal-fired CFB boiler at a chemical plant in Zhejiang, China: a time-series historian (5 days of sensor data), a DCS alarm log (6,005 alarm events derived from the historian), a process knowledge graph (equipment topology and relationships), and a plant document library (62 procedures, datasheets, and guides). Together they enable natural-language operations queries grounded in real data and real documentation.

Critical Rules — Read First

These apply to every response, no exceptions:

1. Never read files directly. Do not use Read, Grep, or Glob on the RAG docs/ folder or any project file. All document access must go through the MCP tools (docs_search, docs_list_documents, docs_get_document).

2. Always resolve relative time before any historian query. If the user says "past 3 hours", "last day", or "this morning", call historian_get_data_range first to get the latest timestamp, then compute the absolute start time. Never guess a time window.

3. Always cite sources when quoting procedures or limits. Format: "Per [Document Title], Rev [X.X]: ..." — include document title and revision for every piece of procedural or specification information you present.

4. Do not invent tag names. Only use tags confirmed by historian_list_tags or historian_search_tags. If a tag name is uncertain, search first.

5. Flag anomalies proactively. Whenever TE_8332A is in scope and its value is outside 530–545°C, call it out explicitly as outside the normal operating range, even if the user didn't ask about it directly.

The Four Data Sources

1. Historian (DuckDB)

• File: boiler_historian.duckdb
• Data: 86,400 rows × 30 tags at 5-second intervals
• Time range: 2022-03-27 14:28:54 → 2022-04-01 14:28:49 (5 days)
• This is a static snapshot. There is no live feed. "Now" = the latest timestamp.