AWS Builder Center
Weekend Creative Challenge: Legendary You

Weekend Creative Challenge: Legendary You

Legendary You turns any person, a coworker, a friend, your mom, into a satirical collectible trading card, generated live with Amazon Nova Pro. Type a name and a few traits, get a holographic card with stats, abilities, rarity, and flavor text.

Vision and what the app does

The people around you are more interesting than any game character. Your coworker who replies-all to a company of 4,000. Your mom who somehow knows you are lying over the phone. Legendary You turns them into collectible trading cards.
You type a name and a few traits, and the app forges a card: a title like "The Reply-All Bard", an element type, HP, four stats (attack, defense, speed, charisma), two special abilities with energy costs, a weakness, a resistance, and a short flavor quote. Dave from Accounting becomes an Epic Spreadsheet-type with a Karaoke Blast that drains 10 HP. Type in "Mom" with "makes the best soup, always wins arguments" and you get The Soup Sage, Mythic rarity, 180 HP.
The output is designed to look like a real card you would pull from a pack. A gold foil frame, a rainbow holographic sheen that drifts across the surface, a floating emoji portrait, rarity stars, and stat bars. The joke lands because the format is so familiar and the subject is so ordinary.
![Legendary You app showing a forged card for Dave, The Midnight Coder: a Caffeine-type card with 120 HP, stat bars, and Caffeine Burst and Night Owl Vision abilities](upload: Screenshot From 2026-08-14 05-04-43.png)
forged card for Dave


How I built it

I built this one to run locally and call the model live, so every card is generated fresh from your exact input.
The core is a prompt that casts the model as a card designer for a satirical collectible game. It has to return one strict JSON object with fixed keys: name, epithet, type, HP, stats, abilities, weakness, resistance, flavor, and even emoji for the portrait and energy costs. Locking that schema means the frontend can render a complex card without guessing. Getting the model to reliably emit emojis inside JSON took a couple of prompt passes, since early versions wrapped the output in code fences or added commentary. A small parser that strips fences and extracts the JSON block made it robust.
The backend is a single Python file using the standard library http server. It serves the page and one endpoint,  /api/generate , which calls Amazon Nova Pro and returns the card. Because the page and the API share an origin, there is no CORS to fight. I capped input length and validated the name to keep requests small.
The most interesting decision was live versus static. My first app in this challenge was deployed publicly, so I stripped its model calls and served pre-generated content to remove any abuse surface. This one is the opposite choice: it stays on my machine, so I can let it call the model on every click without exposing an open endpoint to the internet. Same model, different tradeoff, driven by where the app runs.
The visual style was a deliberate reset from my first project. That one was a dark pitch-deck with serif type. This is bright and playful: blocky display lettering, rounded body text, pastel gradients, and a foil card with an animated holographic layer built entirely in CSS. No image generation, just layout and color doing the work.
Every card comes out different because the traits drive everything, the type, the rarity, the abilities, even the emoji portrait.
Mary

Bob


AWS services and architecture

  • Amazon Bedrock (Nova Pro) generates each card as structured JSON through the Converse API.
Browser (local page)
| POST /api/generate { name, traits }
v
Local Python server (http.server + boto3)
| Converse API
v
Amazon Bedrock: Nova Pro
|
v
strict JSON -> rendered holographic card
The whole thing is two files: a server and a page. Nova Pro does the writing, CSS does the styling, and nothing else is needed.

What I learned

Constraining a model to a rich, fixed JSON schema is what makes a generative app feel designed instead of random. The prompt is the contract.
I also learned that "where does it run" changes the whole security posture. A public app wants no live model call. A local app can call freely because there is no open door. Same code, different deployment, opposite decisions.
On the model side, Nova Pro handled emojis and playful tone well, and the Converse API kept the integration to a few lines. One gotcha worth noting for others: some newer models require an inference profile instead of an on-demand model ID, so check that before wiring a call.

Link

Source code (public repo):  htt Leg
Run it locally with  python3 server.py , open  http://localhost:8000 , and forge a card for someone you know.
Any opinions in this article are those of the individual author and may not reflect the opinions of AWS.
Enjoyed reading this content? Let the author know!

Your likes, comments, shares, and saves help creators reach more builders.

Loading recommendations

Loading article