Brand Artefacts Are Not AI Instructions

The constraint

This was scoped as a lightweight client value-add, not a product build. Fast and cheap were the terms, and they were the interesting part. The question was not what a reusable AI workflow could look like with unlimited runway. It was how much genuine capability could be reached inside a value-add budget without shipping something that would embarrass anyone.

That constraint drove every decision below.

The problem

A university client asked whether their design system could drive AI-generated presentations. The honest answer was not yet, because a design system is a set of rules for humans and a model cannot read it as instructions. Passing a model a link to a Figma file produces confident output that is subtly wrong in ways only someone who knows the brand can see.

So the work was translation, not prompting.

The orchestration

Three tools, three different jobs. Choosing them was the design work, not an implementation detail.

Copilot 365 gathered company tenant information: data the other tools could not reach because it lives inside the Microsoft 365 tenant.

Claude Code built the skill itself: an agentic build of a packaged, testable artefact.

Claude ran the workflow: the target surface for the deliverable.

The output was deliberately built to work in both chat agents and AI code workflows. One artefact, two surfaces, because the people who would use it were split across both.

What the work actually was

A design system is written for humans who already understand design. It assumes context a model does not have: why a rule exists, when it bends, what matters when two rules conflict. So the work was interpreting and encoding: visual rules, tone and narrative structure, slide patterns and format constraints, converted into explicit reusable behaviour a model could actually operate on.

Encoded into the skill as operating behaviour: interview first, generate second (most failure modes in AI-generated documents are requirements failures wearing a formatting costume). Draft in markdown first (cheap to review before committing to a heavier format). Evidence controls and source traceability (claims carry their sources). QA checks built into the workflow, not bolted on afterwards.

Where design and development stopped being separate

The skill only worked because both halves were present.

Reading the design system needed a designer: knowing which rules were load-bearing and which were conventions, what “on brand” means when the guidelines are silent, where a model would produce something technically compliant and visibly wrong.

Building it needed a developer: packaging, testing, iterating on real output, making it work across two runtimes.

Either skill alone produces a worse artefact. A designer without the build skills writes a document about what the AI should do. A developer without the design judgement builds a working pipeline that generates confidently off-brand material.

Then I taught it

The practice manager asked me to present the method to my peers. That mattered more than the artefact. A skill that works for one deliverable is a tool. A method other consultants can apply to their own clients is capability. The session was framed as a consulting example rather than a technical demo, because the transferable part was never the code.

Proof tile

Tool chosen: Claude, via an agentic build in Claude Code. Why: the deliverable was a packaged, testable artefact that needed to run in two environments, which suited an agentic build over interactive prompting. What it got wrong: the delivered deck’s speaker notes shipped with unresolved generation artefacts still in them. Raw internal search-turn citations left where sources should have been. It read as finished. It passed a glance. Nobody in the room noticed. How experience caught it: by reading the notes as a reviewer rather than as the author, on the assumption that plausible-looking output is where the errors hide. Which is exactly what the skill’s own evidence controls were built to enforce, and exactly the check that was skipped on the artefact about the skill.

Hard Skills

Claude, Claude Code, Copilot 365, skill architecture, design system translation, AI workflow design

Soft Skills

AI orchestration and tool selection, design system interpretation, consulting delivery under constraint, method transfer and peer enablement

Note: The source material for this engagement is confidential. No client-identifying information, colleague names, proprietary models or deck content is reproduced here.