KT Cloud: The Evolution from Personas to YAML Harnesses for AI-Generated Code Governance

The KT Cloud Group Business Planning team found that while the adoption of AI coding tools accelerated code generation speed, a bottleneck occurred at the 'verification' stage for checking compliance with team standards. In particular, AI confidently generated code that subtly violated naming conventions or required tags, exposing the limitations of the existing review system that relied on human attention.

Three Stages of Evolution in the Verification System

The three stages the team went through to solve governance issues are as follows:

1. Persona Definition: A method of directly inputting team standards and roles into the prompt. While initially effective, a dilution phenomenon occurred where the AI lost context as conversations grew longer.
2. SKILL.md File Creation: Team infrastructure writing standards, naming conventions, and prohibitions were separated into a Markdown file for the AI to reference. This made management easier, but the structure still required 'trusting' that the AI would follow the file.
3. YAML-Based Harness Verification: Instead of trusting AI compliance, the system enforced rules so that violations could not pass. Resource naming patterns, required tags, and security policies were codified in YAML and automatically verified before GitHub PR merges.

Separation of Automation and Human Roles

Harness applies the same guardrails to both AI-generated and human-written code, ensuring team standards live explicitly within the code. This shifted reviewers' energy from simple checklist verification to judging architectural direction. Final merge decisions are always made by humans, while automation serves to clearly separate what machines can determine from what requires advanced human judgment.