I started MCMDS by documenting everything I knew about my own taste. Color, type, spacing, imagery, product, print. I wanted an agent to have enough material to make the same decisions I would.
Instead, I made a skill that could explain my design system better than it could use it.
The structure was right: universal rules in DESIGN.md, judgment in BRAND.md, production constraints by channel, and reusable assets for output.
The mistake was requiring all of it. More guidance produced more visible design: panels, labels, metadata, technical decoration. It knew the ingredients. It did not know when to stop cooking (sorry).
The first pivot was from copying a visual vocabulary to naming the judgment inside it. Dark fields and document grids mattered less than the behavior beneath them: show evidence, use contrast decisively, let one thing lead. The technical signal layer became conditional. A proposal does not need to cosplay as a control panel because I like exposed electronics.
Scope caused the second pivot. I had treated it as a product reference when the useful part was restraint: one quiet field, compact type, direct content, controls only when needed.
MCMDS became less about applying traits and more about removing defaults. The top-level skill now reads the universal rules first. Taste, creative direction, and channel guidance enter only when the work actually needs them.
The example changed more than another hundred rules would have. An asset review needs the work, the current selection, and the next action. Everything else can wait.
The main lesson is that specificity is not volume. “Use restrained typography” is vague. “Use size as the last hierarchy lever” changes a decision. References work the same way: useful as evidence, dangerous as templates.
There is still a contradiction. Codifying taste makes it repeatable, and repetition is how taste becomes wallpaper. I want the skill to recognize my decisions without freezing them.
Sometimes that means deleting the rule.