What problem does it solve? Changes to mj-agent in-source canonical files (SKILL.md bodies and system.md prompts) risk silent failures like wrong answers, hallucinations, or business drift, and this Skill produces a reviewable EVAL baseline design draft before any evaluation framework exists. ## Core Features & Use Cases - EVAL Design Drafting: Fills the TEMPLATE_EVAL.md 8-section template with eval_kind classification (outcome/trajectory/component/integration), dataset structure, judge selection, and regression thresholds. - Reverse-Scan of Existing Evaluations: Greps target frontmatter eval_references, docs/evaluation/, and tests/eval/ fixtures to avoid duplicate EVAL designs. - Read-Only HITL Workflow: Produces a proposed EVAL document plus structured HITL questions for Domain Expert and Prompt Engineer review without writing to docs/evaluation/ or running pytest. - Use Case: After a B-flavor change to the biz-domain-context SKILL.md, invoke this Skill to draft an outcome-based EVAL with red-line cases covering R1/R2 data boundary rules, then hand the accepted draft to /mj-agent-doc-author for writing. ## Quick Start Ask the agent to propose an EVAL baseline for the biz-domain-context skill and review the drafted TEMPLATE_EVAL document with the HITL questions.
