agent-instructions-evaluator skill evaluates an agent’s instructions for operational achievability and runtime reliability in production environments. It focuses on runtime viability rather than writing style, diagnosing risks such as hidden state, conflicting rules, ambiguous scope, and brittle phrasing.
Source location: skills/agent-instructions-evaluator/SKILL.md in the ADK repository.
Overview
Most agent execution failures in production stem from ambiguous or conflicting instructions rather than model limitations. Theagent-instructions-evaluator skill acts as an automated reviewer, inspecting agent YAML definitions, system prompts, and tool descriptions to surface hidden failure modes before deployment.
Best for: Verifying agent reliability, testing prompt robustness, refactoring complex agent instructions, and CI/CD quality gates.
The 5 evaluation dimensions
The skill scores instructions on a standardized 0–5 scale across five key dimensions:Deterministic signal analysis
In addition to qualitative analysis, the skill measures concrete instruction complexity metrics:- Total instruction line and token count
- Count of nested conditional branches (
if/else,when) - Number of negative constraints (
do not,never,avoid) - Explicit tool reference alignment with bound tools in agent YAML
- Ambiguous trigger phrases
Structured output report
The skill outputs an evaluation report containing:- Overall Readiness Rating: Production Ready / Needs Refinement / High Risk
- Dimension Scorecard: 0–5 rating with strengths and risk factors for each dimension
- Prioritized Findings: High/Medium/Low severity issues with exact line citations
- Actionable Recommendations: Concrete, rewritten instruction examples that resolve identified risks
- Machine-Readable JSON: Structured summary block suitable for automated CI/CD evaluation gates
Activation
In Bob, activate the skill by pasting an agent YAML or instruction prompt:Next steps
- Refactor the flagged instructions using the recommended rewrites.
- Re-run the evaluator to confirm improved dimension scores.
- Test end-to-end agent interactions using the Evaluation Framework.