Many engineering teams give AI coding agents instructions that sound perfectly clear:
Follow existing patterns.
Do not introduce new abstractions.
Use the established architecture. Follow project standards.
A human developer can usually interpret those instructions because the words are only part of the communication. The developer also brings professional experience, familiarity with the codebase, conversations with teammates, code-review feedback, and an understanding of how the organization makes engineering decisions.
The human fills in the missing context.
An AI coding agent does not necessarily make the same expansion. It may recognize the existing pattern in one layer while overlooking another place where the same principle should apply. It may reproduce the visible structure of a nearby implementation without understanding why that structure exists.
This does not mean general engineering principles are useless. It means they are incomplete when they stand alone.
A useful instruction must route the agent toward authoritative context. It should identify which architecture guide, repository standard, testing guidance, data-access convention, security requirement, or implementation example the agent must review before beginning the task.
The goal is not to copy every rule into every prompt. That creates duplication, contradiction, and drift. The goal is to maintain canonical engineering documentation and explicitly direct agents toward the relevant sources.
This is also why some criticism of AI-generated software misses part of the problem.
Critics are right to object to insecure code, fragile systems, unnecessary abstractions, inconsistent implementations, and applications that nobody can maintain. But many of those failures are not proof that AI has no useful role in engineering. They are signs that organizations are using agents without adapting how engineering knowledge is communicated.
Experienced engineers often carry large amounts of organizational knowledge implicitly. They know which patterns are intentional, which compromises are temporary, which architectural boundaries matter, and which production lessons must never be forgotten.
AI systems do not inherit that cultural understanding automatically.
Organizations have to digitize it.
That means turning tribal knowledge into explicit organizational intelligence that can guide both humans and AI systems. It means preserving not only what a codebase does, but why it works the way it does and which decisions should not be casually reversed.
The teams that do this well will not merely generate code faster. They will make the quality standards of experienced engineers available at the point where AI-assisted work actually occurs.
I explore the complete argument in:
Originally published on bdonaldharris.com.