Eval Authoring Tool Agent Skill

Tool Eval Authoring

Author Vally evals that verify one MCP tool is selected correctly for a given prompt, under evals/tools/. All the actual guidance — naming convention, placement, per-category requirements, glossary, grading patterns, grader catalog, anti-patterns, and worked examples — lives in one shared place so it never drifts across the three eval-authoring skills: the repository-local eval authoring guide at .github/skills/eval-authoring/README.md. This skill exists to route you there with tool-eval context already loaded, not to duplicate it.

Triggers

USE FOR: write a tool eval, add prompt-to-tool coverage, test MCP tool selection, add tool catalog eval, create single-tool scenario, harden tool-call grader

WHEN: "write a tool eval", "add prompt-to-tool coverage", "test MCP tool selection", "add tool catalog eval", "create single-tool scenario", "harden tool-call grader"

DO NOT USE FOR: skill routing or capability evals (use eval-authoring-skill), multi-tool, multi-turn, or live scenarios (use eval-authoring-workflow)

Steps

• Read the eval authoring guide (.github/skills/eval-authoring/README.md) Step 0 to find this repo's vallyRoot/evalGlobs for the tool tier, then the guide's "Tool" column throughout (naming, requirements, worked example).
• Confirm the scenario expects one primary tool. If success requires orchestration, conversation state, or live services, use eval-authoring-workflow instead.
• Add stimuli to the matching prompt-to-tool-<area>.eval.yaml; create a separate file only when it needs fixtures or outcome grading.
• Use realistic, concrete prompts with multiple natural phrasings and collision cases that disallow the nearest competing tool. Make tool-calls the primary signal; avoid the anti-patterns in the guide (vacuous keyword-only grading, missing scoring.threshold).
• Keep the unit tier hermetic (environment: azsdk-mcp-mock, tags.tier: unit, established area tag). Avoid git worktrees and production writes.
• Validate locally per the guide's "Running evals locally" section, building the mock MCP first. Do not finish or open a PR until it passes; inspect recorded tool names and arguments on failure.

Rules

• Use exact MCP tool names and assert forbidden alternatives where ambiguity exists.
• Do not force a tool through unnatural prompt instructions unless direct invocation is the contract being tested.
• Keep fixture paths relative to the eval file and fixture data minimal and non-secret.
• Preserve existing namespace coverage and add new stimuli instead of duplicating files.

References

• Eval authoring guide: .github/skills/eval-authoring/README.md — naming, placement, requirements, glossary, graders, anti-patterns, worked examples, local commands. Shared with eval-authoring-skill/eval-authoring-workflow — update it there, not per-skill.
• Repository-local eval README/configuration discovered in Step 0 (if present)