Complete Series · 4 essays · 2025

Instruction Governance

What changes when AI instructions stop being prompts and become production assets?

When an AI system produces a bad answer, the model is an obvious place to look.

But sometimes the model followed our instructions perfectly. The instructions were the problem.

They were ambiguous. They encoded the wrong assumptions. They missed an edge case. They reflected one person's idea of “good.” Or they changed without anyone being able to say what behavior was supposed to remain stable.

Instruction Governance is a four-part Series about what happens when those instructions become consequential enough that prompt engineering alone is no longer sufficient.

The underlying shift is from prompts as implementation details to instructions as governed product assets.

Start with Part 1 ↓

The sequence

From defining good to scaling control with risk

Part 1Define good

Stop Blaming the Model. Start Evaluating Your Instructions.

The Series starts with a problem hidden inside many AI evaluations: teams ask whether the model produced a good answer before asking whether they gave it a good instruction.

That separates Instruction Fidelity from Model Capability, with Data Context as a third variable. It then asks who gets to define “good,” leading to the Governance Triad: Product owns intent, domain experts own specialized correctness and risk, and Engineering owns the system and its automation.

Read Part 1 →

Part 2Encode & test

From Vibes to a Golden Set

Once experts can identify what is wrong, the next challenge is turning qualitative feedback into something a system can repeatedly test.

Real failures are clustered into recurring failure modes. Representative cases become a Golden Set: happy paths, known edge cases and red lines that preserve what the organization has learned about acceptable behavior. Evaluation becomes a reusable regression system rather than a one-off prompt-debugging exercise.

Read Part 2 →

Part 3Govern at scale

From One Team to Enterprise Governance

A Golden Set can work well for one feature. The harder problem appears when dozens of teams each have their own AI systems, instructions and definitions of acceptable behavior.

The Series develops three pillars: the People, the Asset and the Mechanism. Governance becomes federated through a shared Constitution for organization-wide rules and local Bylaws for product-specific behavior.

Read Part 3 →

Part 4Scale by risk

Governance Is a Dial, Not a Switch

The final problem is practical: how do you introduce rigor without turning AI development into bureaucracy?

The answer is to scale governance with both maturity and risk through Crawl, Walk, Run. A low-risk internal tool and a regulated consequential workflow should not carry the same control burden.

Governance should increase with consequence, not simply with organizational size.

Read Part 4 →

From prompt engineering to organizational control

The important unit is no longer the prompt. It is the behavior the organization intends the AI system to preserve, together with the evidence, ownership and tests that make that behavior inspectable.

Prompt engineering asks whether we can make the model behave. Instruction Governance asks whether we can define, test and maintain the behavior we are willing to stand behind.