Agentic Content Review · Mastercard

Agentic Content Review · Mastercard

Role: Content Systems and AI Enablement Consultant When: October to November 2026 Status: Proof of concept, delivered as a versioned bundle. Light testing. Not in production Scope: AI review of interface copy for MDES Manager, against the team's written content standards, in Figma's AI agent and Claude Code

⚠️

What this is and is not. I designed and built this with the product's design and content team as a proof of concept. It has had light testing. It is not live in production and I claim no measured saving. The client's content designer contributed code to the orchestrator.

The Problem

Interface copy on an enterprise product comes from many hands and is reviewed screen by screen against written standards. A reviewer applies the same dozens of rules by hand, every release. Some rules are mechanical, such as casing and full stops. Others need judgment, such as tone in an error message.

The question was how to hand part of that review to an AI agent without building a confident tool that flags the wrong things at scale.

What I Built

PartWhat it is
Audit166 product screens and 28 message captures reviewed. Each finding became a rule, traced to a source guideline or marked as a project decision
SkillsA library of AI review skills. 27 delivered in a versioned bundle with a change log, about 75 defined. Each has a Figma agent version and a Claude Code version
OrchestratorThe pipeline that runs the skills on a screen and writes one result
EvalsA rubric of six checks per screen, a run folder per run, and traces. The harness repeats the client's own tests flow by flow

I also ran the workshop that set the decisions the rules follow.

How the Orchestrator Works

  1. Read the screen once. One step works out what kind of screen it is and who uses it. Every skill gets the same brief
  2. Detect in parallel. Every detector runs on the same frozen original text. No skill reads another skill's edits
  3. Settle conflicts in one place. A single precedence table decides when two rules disagree
  4. Write once. One write per screen replaces a chain of patches
  5. Run mechanical rules in code. Casing, full stops, spelling, dates and times never reach the model. A deterministic pass applies them afterwards
  6. Guard the facts twice. A fact ledger, then a token comparison against the original: numbers, codes, units, placeholders, variables, URLs and fixed names
  7. Grade separately. A grader that has not seen the writer's work scores the final copy against the rubric. A check that passed on the original and fails on the result is a regression

Decisions That Shaped It

Sort the rules by how reliably they can be applied. Simple rules have one test and pass or fail. Hybrid rules are a pattern plus a decision. Probabilistic rules need language judgment. Code does what code can check. A model only does judgment.

One rule family per skill, with its evidence. Every rule is quoted from a guideline or marked as the project's own decision. If the guideline is silent, the skill says so.

Let the platform limits shape the design. Figma's agent takes one Markdown file per skill, with no scripts and no folders. So the skills carry no code and load shared knowledge by name.

Record judgment calls once. A term register keeps each call the orchestrator makes on a board, so the next run makes the same call.

Test the way the client tests. The evals mirror the content designer's own method, so the two sets of results can be read side by side.

Honest Limits

  • It is a proof of concept. Testing was light, and on constructed test strings
  • Not every skill was re-run after the last rewrite
  • Token savings from the orchestrator are not measured
  • Some rulings are still open, such as full stops on single sentences and contractions
  • I do not know which model sits behind Figma's agent. The rules are plain Markdown so they survive a change of model

Why It Matters

This is solution design applied to AI. The decisions are the ones in any architecture review: where deterministic code ends and a model begins, what the failure modes are, how it is tested and who owns each rule. Most people who use an LLM can produce an output. Fewer can show the guard on facts, the regression rule and the separate grader.

It builds on my earlier MDES Manager work and the practice described under AI Enablement.


→ Mastercard MDES for the wider engagement → AI Enablement for how I work with LLM tooling → Contact to discuss a role