The architecture I designed, then cut to what could hold it

The situation

I designed a layered architecture to govern generated copy. The base layer held the fundamental writing rules, the brand foundation, tone of voice, grammatical nuances, and the language the brand never uses. Above it sat a layer per content type, then campaign style, then market compliance, then voice per product.

Five layers, each one narrowing the last. On paper it was the right answer. Every layer existed because a real distinction existed, and I could defend each one individually.

The diagnosis

Then I put it in front of the engineers who would have to run it and asked what would break. Two problems came out of that room. One was in the system. The other was in me.

The implementing system could not hold that depth without paying for it in variance. Each additional layer was another instruction competing for the model's attention, and past a certain depth the output got less predictable. It inverts the intuition the design was built on, producing more randomness instead of constraining it. The whole purpose of standardising a brand's output is to reduce variance, and the proposed architecture does the opposite, even in the name of control.

The second one is a pattern rather than an incident. I build systems more sophisticated than the people who have to run them can adopt. That failure shows up in one of two ways, and the second is worse. Either nobody can catch up, so the system sits unused while its author explains it; or people follow it exactly without understanding it, which produces compliance instead of judgement and hides the failure until something needs a decision the rules do not cover.

The limit was information, and it came from the only people in a position to give it. The paper argument for the design still held. The system it had to run on did not.

The mechanism

I cut five layers to two, and the rules that came out of doing it are the transferable part.

  1. Depth is bounded by what the implementing system can hold without adding variance. Test the depth, do not assume it. The ceiling is empirical and it moves as the tooling changes.

  2. The right depth is set by whoever maintains it, not by whoever designs it. If the maintainer cannot hold the whole model in their head, the model is wrong for that team regardless of how correct it looks.

  3. Every layer is another place for drift. Five layers means five opportunities for two rules to quietly disagree, and no one notices until the output is strange in a way nobody can source.

  4. Collapse by absorption, not deletion. The distinctions did not disappear. Compliance and product voice moved into the layers that survived, as rules rather than as levels.

Where the line sits

Cutting the architecture did not mean standardising less. It meant standardising the right things. I am not arguing against AI in creative work, only against adopting it without drawing the line first.

Standardise what should be identical. Terminology, banned words, mechanical formatting, the voice of a given channel, the structure of an intake, the bands of an evaluation. Consistency here is not a compromise of craft. It is what a brand is, and it is what commercial work owes its reader.

Protect what should differ. The concept, the perspective, the uncomfortable idea, and the judgment about whether the brief asked the right question. No layer count protects these. Only a decision does.

What transfers

Standardisation has an optimum depth. Most teams building this kind of governance stop at "more rules, more control" and never find the point where their curve turns. Find it with your implementers in the room, before you have built five layers and have to argue yourself down from them.

The habit I have kept from this: before shipping any system, translate it down. If someone new to the field cannot follow it without me in the room, it is not finished, however correct it is.

One thing I will not dress up. Standardising carefully still reduces the number of people a function needs, and the practitioners carry that risk whether the adoption is thoughtful or reckless. Doing it well changes who stays and what they get to do with their time. It does not change the arithmetic. Anyone selling this work should say so out loud.

Previous
Previous

Optimised to a Farm: On AI Adoption in a Creative Function

Next
Next

We had criteria, and we still could not agree