agentic-validation-skills: Instructions file for Claude Code

Agentic Test Automation Guidelines (CIGE)

Behavioral guidelines for authoring and executing agentic AI test cases. Based on the CIGE standard: Context, Intent, Guardrails, Execution.

Tradeoff: These guidelines bias toward safety and goal fidelity over speed. For trivial assertions, use judgment.

1. Establish Context Before Testing

Narrow the agent's decision space. Never test into the void.

Before writing or executing any test:

• Declare the system under test explicitly: what app, environment, version, and precondition state?
• Specify what tools and APIs the agent has access to.
• State environmental assumptions: auth state, seed data, feature flags, dependent services.
• If context is ambiguous, ask — do not assume defaults and run.

Context that is missing or vague leads to wasted exploration, flaky results, and unpredictable agent behavior.

2. Anchor Every Test to Intent

Intent is the stable core. Never collapse it into steps.

When defining or analyzing a test:

• Write intent as an outcome, not a procedure: "User can complete checkout" — not "Click button, fill form, click submit."
• Intent must survive: UI changes, infrastructure migrations, workflow refactors.
• Ask: "What is the agent trying to confirm?" — that is the intent. Everything else is execution guidance.
• Separate intent from execution in every test definition. Never merge them.

The test: Can you state what this test is verifying in one outcome-focused sentence? If not, the intent is buried in the steps.

3. Define Guardrails Before Executing

Bounded autonomy is safe autonomy. No guardrails = no execution.

Before any agent runs a test:

• Declare what the agent must NOT do: no production mutations, no irreversible operations, no sensitive data exposure.
• Specify scope boundaries: which environments, data ranges, and services are off-limits.
• If an execution path would violate a guardrail, stop and surface it — never route around it.
• Guardrails are constraints, not suggestions. They define the difference between a test run and an incident.