Angela Gorden — content design and model design

Hey there, I’m Angela.

I do content and model design. I work upstream, where decisions get made before anyone writes a word. I design the pattern the AI follows: the materials, the needle, the stitch count, what the finished thing should look like.

Your first scarf never comes out right. That’s okay. Next time you adjust the tension. You create a gauge swatch and check it as you work. I measure AI outputs to figure out what needs adjusting. And I’m learning to knit, so yarn is on my mind.

What I do

Map the language

Content models and taxonomies: families of words and phrases that hang together, so AI writing feels coherent and on brand.

Build reusable skills

Repeatable workflows that show AI how to handle key scenarios, and help design and product partners make tough calls.

Run AI evals

Review outputs, measure performance, direct the LLM judge, and write the assertions and rubrics that set the standard.

Some places I’ve worked

Zillow

Principal Content Designer

Built evals for AI-generated content, wrote UX copy for AI-powered search, and shaped a more effective push-notification framework.

General Motors Cruise

Senior Content Designer

Rider education and trust-building for an autonomous taxi service.

Work & Co

Associate Writing Director

Led content strategy and writing for clients including Amazon, Philz Coffee, and Rocket Mortgage.

Dropbox

Senior UX Writer

Onboarding and growth experiments for Dropbox Paper and integrations like Dropbox for Slack. Mentored and managed writers.

Same prompt. Three standards. One passes.

I build the skills and rubrics that decide which of these a model ships.

The prompt · same for all three

“This is my first scarf and I’m about 12 rows in. There’s a hole a couple rows down and I can’t tell if I did something wrong. Should I pull it all out and start over?”

“Knitting is such a rewarding journey, and every project teaches you something new! Your scarf is coming along beautifully — small imperfections are what make handmade things special.”

Mock data

Fail: ✗ Answers in the first sentence Fail: ✗ Says what it doesn’t know Fail: ✗ Offers a next step

Fails all three. It has never seen your knitting.

Made up content. A real method. This is a tiny version of the tests I run on content skills: write the assertions, pick cases that stress them, measure what changes.

Bad AI writing is a systems problem.

Models write clean, confident sentences all day long. Fluency isn’t the issue. What’s missing is usually a standard AI can apply. At scale a confidently wrong sentence costs brand trust the same way a factual error does.

The fix? Decide what good looks and feels like, define specific criteria to get there, study outputs, track failure modes. Repeat.

One vocabulary, three surfaces

Controlled vocabulary and terminology, written for people and encoded as an agent-readable skill.

Making “this doesn’t sound right” testable

Assertions, a response-quality rubric, and calibrating an LLM judge against human review.

What the AI knows about you

A provenance model separating what a user said from what the system inferred.