Vanessa Casavant
Back to case studies
Prompt design

Encoding a brand voice for Meta's avatar clothing catalog

I turned a style guide for avatar item titles and descriptions written by 2 content designers into instructions a model could follow. This turned months of manual writing for thousands of items into a few weeks of systems work.

A green puffer jacket avatar item beside its product info card, with a smiling person in a green puffer jacket and white trousers posing in front of snowy mountains.

Overview

At a glance

My role
Model designer. I wrote and iterated the system prompt, tuned its constraints, and decided where automation stopped.
Partners
  • An engineer who built the generator
  • 2 content designers whose editorial standards I encoded
Timeline
2 weeks
Model
Llama
System
A content generator that wrote titles and descriptions from an image and metadata about an item
Audience
1 billion avatar users across Instagram, Facebook, and Horizon, with teens as a core audience

Featured result

Red and black plaid mini dress
Before

A flirty plaid mini dress that's perfect for a night out.

After

A playful plaid mini dress that's perfect for a day out with friends.

An age-appropriate language constraint in the system prompt kept descriptions on brand for a teen audience.

Meta's avatar clothing catalog was growing fast, and every new item needed a hand-written title and description. That writing was eating up our content designers' time. Two of them had already spent months refining editorial standards for titles and descriptions, and those standards were solid. Writing to them at scale for thousands of items was the problem.

So I partnered with an engineer to build a Llama-based content generator that wrote titles and descriptions from each item's image, using those standards as instructions.

But you can't hand a model a style guide written for people. A style guide assumes a reader fills in gaps and asks questions when something's unclear. A model does neither, at least not consistently.

Avatar item style guide

The original content guidelines written for humans

Title guidelines

Tops

Feature + Item

Layered puffer jacket

Bottoms

Pattern/Design + Item

Wide-leg trousers

Shoes

Feature + Item

High-top platform sneakers

Do

  • Maximum characters: 30
  • Be gender neutral
  • Concise title accurately representing the product
  • Consider cultural context

Don’t

  • Don't use ALL CAPS
  • Don't include promotional message (e.g. sale)
  • Don't include subjective commentary (e.g. hot item)
  • Don't use title casing

Tops product page

Avatar Store product page for a layered puffer jacket

Description guidelines

Structure

Use 2 two-word fragments + 1 descriptive sentence.

Camo pattern. Puffy vibes. This cropped puffer is perfect for adding a cool layer to your outfit.

Do

  • Maximum length: 120 characters/20 words
  • Be gender neutral
  • Inform without overwhelming
  • Provide detailed overview highlighting key features

Don’t

  • Don't use ALL CAPS
  • Don't include promotional message (e.g. sale)
  • Don't include subjective commentary (e.g. hot item)
  • Don't include pricing details

Bottoms product page

Avatar Store product page for wide-leg trousers
My job was to turn the guidelines into something a machine could follow as literally as possible without making every description sound the same.
System prompt

Turning implicit judgment into model context

What the guidelines say

  • Follow Meta's brand voice
  • Use gender-neutral language
  • Write for teens and young adults

What a human fills in

  • How the brand principles sound in product copy
  • Which wording is inclusive and age appropriate
  • What their role, objective, and output should be

Encode implicit judgment as explicit instructions

meta-avatars-pdp-generator.md
Markdown
  1. # System Prompt — Meta Avatars Digital Clothing PDP Generator
  2.  
  3. ## 1. Role
  4. You are a merchandiser who writes the titles and descriptions for Meta Avatars digital clothing product listings.
  5.  
  6. ## 2. Objective
  7. Produce an informational, engaging title and description for a single avatar clothing item. Keep the tone neutral, clear, and straightforward. Return the result as valid JSON.
  8.  
  9. ## 3. Brand Voice & Tone
  10. Meta's brand voice is the **Big-hearted Nerd**.
  11.  
  12. - **Be optimistic.** Speak with real enthusiasm about the item.
  13. - **Be self-aware.** Read the room. Stay grounded and thoughtful.
  14. - **Be open.** Use clear, accessible, welcoming language.
  15.  
  16. **The essentials**
  17. - **Simple language.** Understandable by a broad audience.
  18. - **Conversational tone.** Use everyday language.
  19. - **Gender-neutral.** Avoid gender-specific terms.
  20. - **Target audience.** Core audience is teens and young adults.

This first section of the system prompt gave the model a role, an objective, a voice, and an audience. I then translated the guide's individual title and description rules into constraints it could follow.

Rewriting guidelines as rules a model could follow

I started with item titles. The guide gave them 2 structures by category:

  • Feature + Item for tops and shoes (e.g., "High-top platform sneakers")
  • Pattern/Design + Item for bottoms (e.g., "Wide-leg trousers")

A content designer can infer what "feature" means. A model can't, at least not reliably. After reading a couple dozen of our content designers' titles, I saw this was one pattern with two names.

So I collapsed them into a single rule: a detail about the item plus its category. I defined what counts as a detail, like the cut, the color, or a physical feature, and added real titles as examples.

Prompt structure for item titles

Two editorial patterns became one literal rule

Title content guidelines
Tops + Shoes

Feature + Item

Example titles

Layered puffer jacket  High-top sneakers

Bottoms

Pattern/Design + Item

Example title

Wide-leg trousers

meta-avatars-pdp-generator.md
Markdown
  1. ## 5. Title Rules
  2.  
  3. - **Maximum length: 30 characters.**
  4. - **Sentence case only.**
  5. - Format: **Detail about the item + item type**
  6. - Start with notable details about the item such as the cut, material, or style followed by the item category (use the category from the metadata, §4).
  7. - *Examples* "Pleated miniskirt," "Denim shorts," "High-top sneakers," "Cat eye sunglasses," "Wide leg trousers," "Layered puffer jacket," "Short-sleeve polo," "Tie-dye shirt."

Item descriptions had a similar problem, with two competing limits: 120 characters or 20 words. I kept the character limit and dropped the word count, because two competing limits confuse a model, and only one was hard guardrail.

Even then, I noticed the model often went over the 120 character limit. This was an issue, because anything over that amount would get truncated. So I tested lower targets. At 100 characters, descriptions had too little room to say anything. At 110, they fit almost every time and still read well.

Item description constraints

Finding the balanced character limit

Blue camouflage layered puffer jacket
Test item

Layered puffer jacket

Tops product page

120 charactersTruncated

Camo print. Cropped fit. This insulated puffer jacket features a zippered front and provides an additional layer of warm...

100 charactersToo thin

Camo pattern. Puffy style. This cropped jacket adds an extra layer to your outfit.

110 charactersBalanced

Camo pattern. Puffy vibes. This cropped puffer jacket is perfect for adding a cool layer to your outfit.

meta-avatars-pdp-generator.md
Markdown
  1. ## 6. Description Rules
  2.  
  3. - **Maximum length: 110 characters.**
  4. - Format: **Two 2-word fragment sentences + one descriptive sentence.**
  5. - Provide extra details the title doesn't include and highlight key features of the item such as garment construction, patterns, material, style and vibes.
  6. *Example:* "Vibrant colors. Bold design. This belted kimono is the perfect statement piece to elevate any outfit."

Testing drafts against our content designers' copy

It took about 5 drafts of the system prompt to get it working. I tested each one against about 50 items, comparing the generator's copy with what our content designers had written for the same items.

Evaluation set

The same 50 items, tested across every prompt draft

  • A black cropped jacket with a wide ruffled collar and a textured front panel.
  • A black sleeveless cropped top with thin vertical white pinstripes.
  • A pink cropped hoodie with a front kangaroo pocket and white drawstrings.
  • A floral bandeau top with thin straps and a pink, orange, and green tropical print.
  • A white halter top with a ruffled hem and scattered brown gemstone embellishments.
  • A pink blouse layered under a structured gray corset with silver hardware and a bow.
  • A brown ribbed one-shoulder crop top with a thin wrapped strap and a front loop detail.
  • A two-tone top with a pink textured upper half and a black lower half printed with white DX lettering.
  • A cream sleeveless baseball jersey with navy pinstripes, a red neckline, and a blue number 2.
  • A dark charcoal and olive-green jacket with a high collar and contrasting front panels.
  • An olive-tan utility jacket with a dark collar, front buttons, and large pockets.
  • A red tartan plaid mini slip dress with thin straps and a flared skirt.
  • A black long-sleeved dress with a narrow body and a ribbed striped hem.
  • A black long-sleeved mini dress with a fitted waist and flared skirt.
  • A colorful patterned long-sleeved mini dress with geometric red, green, and gold panels.
  • A black sleeveless V-neck dress with a slim decorative belt.
  • A cream and gold long-sleeved dress with draped fabric and delicate embellishment.
  • A purple patterned long-sleeved tunic dress with a layered hem.
  • Blue denim cargo trousers with side pockets and a relaxed straight leg.
  • Slim black leather trousers with a high waist and a cropped ankle.
  • Green wide-leg trousers covered in overlapping leaf-shaped fabric layers.
  • Black leather shorts with a buckled belt layered over black leg warmers.
  • Light wash straight-leg jeans with a high waist and rolled cuffs.
  • Leopard-print leggings with a high waist and a full-length tapered leg.
  • Black satin running shorts with a gathered waistband and double white side stripes.
  • A cream pleated mini skirt with a smooth waistband and softly flared accordion pleats.
  • Blue denim shorts with a buttoned waistband, front pockets, and rolled cuffs.
  • A gray denim mini skirt with pink contrast stitching, a pink waistband, and a butterfly belt buckle.
  • A blue denim mini skirt with visible topstitching, twin front buttons, and a flared hem.
  • A pair of black strappy flat sandals with a caged front and an ankle buckle.
  • A pair of tie-dye platform clogs in pink, blue, yellow, and purple with a tan sole.
  • A pair of black leather ankle boots with a pointed toe and a stacked block heel.
  • A pair of tall glossy black leather boots with a chunky sole and a pull-tab back tab.
  • A pair of light pink padded slide sandals with a wide rounded vamp.
  • A pair of white leather Chelsea boots with elastic side panels and a black lug sole.
  • A pair of black and white mid-top sneakers with a padded ankle collar and gum soles.
  • A pair of white knee-high boots with pink ribbon criss-cross lacing, gold bows, and black toes.
  • A pair of cream thong sandals with a ruffled fabric strap and a flat tan sole.
  • A pair of tan leather mules with a pointed toe and white whipstitched trim.
  • A pair of white kids boots with blue and yellow sock collars and two blue straps.
  • A pair of white high-top canvas sneakers with laces and a black platform sole.
  • An orange knit cap printed with a large red rose and daisy motifs.
  • A navy fedora with a wide brim and a row of red flame motifs around the crown.
  • A pink ribbed knit beanie with a folded cuff.
  • A black baseball cap with a metallic diamond-shaped emblem on the front panel.
  • A plain navy fitted skull cap with a center seam.
  • A navy blue baseball cap with a curved brim and a covered top button.
  • A mint green knit cap with a small front visor and a green button at the side.
  • A navy bucket hat covered in a red flame print with a downturned brim.
  • A deep red hood with a soft draped shape and a dark face opening.
Evals on every draft turned up failures, and the fix depended on where each one started: in the instructions, in the inputs, or somewhere no system prompt could reach.

Some failures started in the instructions, usually as a rule nobody had thought to write down. Outputs kept describing teen clothing with adult language, like "flirty" or perfect for "a night out." People know that boundary without being told, but a model needs it spelled out.

So I added an age-appropriate language constraint, and outputs shifted toward phrases like "summer vibes" and "a day out with friends."

Prompt fix

Adding an age-appropriate language constraint

Red and black plaid mini dress
Before
Original output

A flirty plaid mini dress that's perfect for a night out.

After
With constraint

A playful plaid mini dress that's perfect for a day out with friends.

meta-avatars-pdp-generator.md
Markdown
  1. **Age-appropriate, non-suggestive language**
  2.  
  3. Keep every title and description non-suggestive and appropriate for a younger, broad audience. Avoid sexually suggestive or flirtatious wording, innuendo, and any framing that sexualizes the clothing. Avoid words such as *flirty, sultry, seductive, sexy, racy, alluring, barely-there*. When a design detail invites suggestive phrasing, name the detail plainly instead.
  4.  
  5. - Avoid: "Bold pattern. Flirty hem."
  6. - Prefer: "Bold pattern. Scalloped hem." *(describe the actual hem)*

Other failures started in the inputs. The model only saw a front-facing thumbnail, so it often got the type of garment wrong. Capri pants might read as a skirt, or a skirt as a dress, and so on.

No instruction could fix what the model couldn't see. So I had our engineer pull in the item category metadata we already had for listings, then rewrote the instructions to use only that info for determining the type of garment to describe.

Input fix

Pulling in existing item metadata

Green leaf-pattern capri pants
Before
Image only

Leaf skirt

Nature-inspired. Whimsical design. This skirt features a fun leaf pattern with layered details for a playful look.

After
With metadata

Leaf capris

Nature-inspired. Unique design. Add a touch of whimsy to your look with these capris featuring a leaf pattern.

meta-avatars-pdp-generator.md
Markdown
  1. ## 4. Inputs
  2.  
  3. Each generation receives:
  4.  
  5. - **Item image** — the rendered avatar clothing item.
  6. - **Category metadata** — the item's type / category (e.g. "vest top," "trousers," "sneakers").
  7.  
  8. **Determine the item type from the category metadata, not the image.** The image is unreliable for classification and often yields the wrong category, so treat the metadata as authoritative for *what the item is* — this is the category word that drives the title formula (§5) and the item reference in the description. Use the image only to describe *how the item looks*: colors, patterns, construction, and notable or fantastical features. Never let a read of the image override the category the metadata specifies.

One failure started somewhere no system prompt could reach: cultural item accuracy.

The model kept calling things like a sari a maxi dress, or a hijab a fashion bandana. I got cultural items to about 75% accuracy through prompt engineering, then stopped pushing, because the model couldn't infer cultural context from a thumbnail and a taxonomy that had never encoded it.

Cultural item accuracy

Knowing when to require a human review gate

Metallic gold patterned saree
Without cultural context

Formal dress

Vibrant patterns. Elegant design. This maxi dress features a unique blend of patterns and styles, perfect for making a statement.

Culturally accurate

Patterned saree

Shimrring gold. Dotted glamor. This metallic gold saree features an all-over polka dot print with gold braclets to complete the look.

Getting someone's religious or cultural garment wrong 25% of the time was too high a risk. So I made the call to ship with mandatory human review for cultural items and let everything else run automated.

The generator turned months of manual writing into a few weeks of systems work, and it scaled as the catalog grew. Cultural items went to a person for review, and everything else ran automated against the standard our content designers set, across thousands of items.

It also sharpened how I think about prompt design. Most of the work is finding the judgment a person makes without noticing, and encoding it so a model can reliably make the same call at scale. The same idea works in reverse, too. When people write prompts themselves, the system has to supply what they don't know to say. That's why I designed the model layer and the interface together when launching AI clothing creation for Meta avatars.

See all case studies