AI for Design Ops

Agents produce fast. I design the conditions to trust what they deliver.

Moving fast with AI is easy. Knowing whether what it produced is correct — that's a different question.

An agent can create a component in 3 seconds. It can also create a component that looks right, compiles, and silently bypasses the design system.

Trust doesn't come from the agent's capabilities. It comes from the system built around it. My work: designing that system.

Trust by task

Understand

Notes, qualitative feedback, quantitative inputs

Medium
High
Medium

Structure

Problem framing, workflows, decision criteria

Medium
Medium
Medium

Explore

Options, scenarios, alternatives, edge cases

High
Medium
High

Prototype

Interactive flows, states, first implementation paths

High
Medium
Medium

Evaluate

Critique, test scenarios, risks, blind spots

Medium
High
Medium

Decide

Trade-offs, product risks, solution choices

Medium
Medium
Low

Document

Specs, memory, guidelines, design system notes

High
High
High

Communicate

Narrative, alignment, presentations, team conversations

Medium
Medium
Low

Produce — without guardrails

Components, tokens, accessibility: no verification

High
Low
Low

Produce — with evidence

Verified components, respected tokens, declared compliance

Medium
High
Medium

Agentic Design System

This is one of the flows showing how my design system is implemented and synchronized between code and Figma, while addressing trust issues: context, evidence, checks and guardrails.

1.0.X

Patch or fix

Design system edited

"Rename the primary color to brand"

Figma

Trigger

Protocol triggered manually or by automated routine

PromptRoutine

Orchestrate

An agent analyzes intent, evaluates risk level and routes to the right lane

skillgovernance.md

Implement

Implementation of the request in code

codetemporary-rfc.mdcomponentName.md

Verify

Certify Figma-code parity

figmacodetestsskillcompliance.md

Approve

Human validates the PR

git

Deliver & memorize

Sanity check and memory update

gitskillmemory.md

Design system infrastructure

Every file has a precise role. Agents read, write and update these files at each step of the flow.

/AGENTS.md
/AGENTS.md
# AGENTS.md

## Role
Design system agent. You contribute to maintaining
and evolving the DS in a governed way.

## Absolute rules
- Read /system/governance.md before any action
- Check /components/component-catalog.md before creating
- Never hardcode token values
- Declare every deviation in a compliance report

## Work cycle
1. Read the context files (/system)
2. Analyse the request and evaluate the risk
3. Run the flow for the matching lane
4. Produce deliverables in /reports
5. Update /system/memory.md

What happens when the agent gets it wrong?

01

What happened

A component implemented without min/max-width and min/max-height tokens, despite them being present in the Figma.

Why

The scan skill didn't target sizing constraint properties, so they were silently ignored.

How I fixed it

scan-figma.md updated with explicit rules to extract all sizing and constraint properties.

02

What happened

The agent created a Button component that already existed at v2.4 in the design system.

Why

The catalog listed component names but not variants; the agent didn't recognise the existing coverage.

How I fixed it

component-catalog.md enriched with variants, states and version for each component.

03

What happened

On a composite page, the agent recoded existing components from scratch instead of reusing DS ones.

Why

Figma layers had inconsistent names, so the agent couldn't connect them to existing code components.

How I fixed it

Layer naming convention + strict recognition rules added to AGENTS.md. Added a naming-check agent.

AI for Design Ops — Quentin Gillon