Instruction file imported from sbroenne/mcp-server-excel (.github/instructions/llm-testing-philosophy.instructions.md). Copyright stays with the author.

LLM evaluations

• Evaluate product discoverability: natural Excel requests, not command/flag tutorials. Fix missing guidance in its canonical source, not by coaching the test agent through the failing step.
• Use build_excel_cli_eval/build_excel_mcp_eval and shared budgets from conftest.py. Normal skill evaluations include the skill; explicitly named tool-description-only or tool-availability experiments may omit/restrict it.
• Shared workflows need equivalent CLI/MCP requests and outcome checks. Transport assertions and documented entry-point-specific experiments can differ.
• Inspect tool calls and workbook outcomes, not just the final prose. Distinguish product failures from auth, Excel, harness, and budget failures. Missing prerequisites may skip in the harness; skip/xfail must not hide product defects.
• Setup and commands live in llm-tests/README.md. Run affected scenarios only: these require Excel and external model access and may incur costs.