AWS Builder Center
Weekend Creative Agent Challenge: Memory Postcard

Weekend Creative Agent Challenge: Memory Postcard

An autonomous creative agent that generates a new AI-illustrated postcard on its own daily schedule, no user prompt required. Built with EventBridge, Lambda, Amazon Bedrock, S3, and Amplify, it picks a fresh theme, decides the scene, and has a finished postcard waiting whenever you return.

Vision & what it does

Memory Postcard started as a simple idea: turn a written memory into a keepsake — a poetic retelling, a title, a quote, and an illustration, composed into something that feels like an actual postcard rather than a raw AI output. That version worked, but it had one limitation shared by almost every creative AI tool: nothing happened until a person opened the app and clicked a button.
For the "Set Your Creative App Free" challenge, the goal was to remove that dependency entirely. Memory Postcard is now an always-on creative agent — it wakes up on its own schedule, makes a creative decision, generates a complete illustrated postcard, and has it waiting the next time someone opens the app. No prompt, no click, no user in the loop at generation time.
The agent doesn't just repeat a template with a new color. It picks a theme — Rainy Day, Snowy Morning, Sunset Escape, Vintage Vibes, City Lights, Monsoon Memories, and others — and that theme changes the actual scene: environment, weather, lighting, objects, and composition, not just a filter over a generic background. It also looks at its own recent output before deciding what to make next, so it doesn't settle into generating the same "Cozy Evening" postcard on repeat.

How I built it

The first version of theme variation was disappointing — different themes were really just different color gradients over the same generic landscape shape. It technically satisfied "themed," but it didn't feel creative. The fix was to stop treating "theme" as a styling parameter and start treating it as a full scene specification: environment, time of day, lighting, foreground/midground/background detail, and artistic treatment, all defined per theme before the image model ever runs. Vintage Vibes, for example, isn't a sepia filter — it's period architecture, vintage signage, and analog-photography characteristics baked into the actual prompt.
A second key decision was separating artwork generation from postcard composition. Image models are unreliable at rendering clean text, so the image model only generates the scene — title, date, and postcard metadata are rendered by the application on top of the artwork. That gave consistent typography without fighting the model for legible text inside the image.
The autonomous trigger itself was the most satisfying piece to get right: moving generation off a user click and onto a scheduled event meant rethinking the Lambda as an independent worker with its own state (recent themes, recent moods) rather than a stateless request handler.

AWS services used / architecture overview

  • Amazon EventBridge Scheduler — triggers the generation workflow automatically, with no manual initiation
  • AWS Lambda — runs the creative-agent logic: theme selection, prompt construction, orchestration
  • Amazon Bedrock (Nova models) — generates the creative content (title, poem, quote) and the illustration itself
  • Amazon S3 — stores the generated artwork and postcard metadata
  • AWS Amplify — hosts the frontend where completed postcards are displayed
Amazon EventBridge Scheduler
│
▼
AWS Lambda
│
Creative Agent Logic
(theme + history-aware)
│
▼
Amazon Bedrock
(content + image generation)
│
▼
Amazon S3
(artwork + metadata)
│
▼
Memory Postcard UI (Amplify)
│
▼
User returns later —
postcard is already waiting

What I learned

The biggest shift was conceptual, not technical: an autonomous app needs its creative logic decoupled from the UI entirely, with its own sense of state and history, rather than living inside a request/response cycle. I also learned that visual diversity has to be engineered into the prompt at the scene level — swapping colors doesn't produce a different world, but swapping environment, lighting, and composition does. Separating artwork generation from typography rendering turned out to be a small architectural decision with an outsized effect on output quality. And working with EventBridge Scheduler for the first time was a good reminder that "autonomous" isn't a UI feature — it's an infrastructure decision.

Link to app or repo

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