Eval Criteria: ai-layer-kit/agents/explorer.md

Created: 2026-05-26 Target: production-lines/digital-talent/templates/ai-layer-kit/agents/explorer.md Purpose: Subagent definition for read-only code/content exploration. Frontmatter description drives parent-agent triggering; body is the subagent's system prompt.

Questions

Triggering effectiveness (frontmatter description)

1. Does the description start with an action verb (what the subagent does), not a noun phrase?
2. Does the description contain at least one explicit "use when …" trigger to guide the parent?
3. Does the description state the explicit non-action ("cannot write" / "read-only") so the parent doesn't mis-call?

Tool / prompt alignment (safety)

4. Is the tools: frontmatter list free of write-capable tools (Write, Edit, NotebookEdit)?
5. Does the body explicitly name Write / Edit / NotebookEdit to forbid them (defense in depth with the toolset)?
6. Does the body define acceptable vs. forbidden Bash usage (since Bash is in the toolset and can mutate state)?

Operational clarity

7. Are the operating rules numbered (so they can be cited by number when iterating)?
8. Does the file include a "when NOT to call" section addressed to the parent agent?
9. Is the expected output format provided as a literal template the parent can pattern-match?

Token economy

10. Is the file under 50 content lines (the file loads on every dispatch)?
11. Does the body avoid restating content already in the frontmatter description (no semantic duplication of the no-write theme beyond the intentional defense-in-depth in Rule 1)?

Specificity

12. Does at least one operating rule contain a numeric threshold or concrete example (avoiding vague "be concise / be careful" language)?