
Weekend Showcase Challenge: Daily Postcard Agent
An always-on AWS agent that paints a new postcard every morning from the day's weather — and, new this weekend, closes the loop by having Bedrock critique its own rendered image rather than just its prompt, so the daily style evolution is grounded in what was actually painted.
Vision and what it does
For the season finale I went back to my strongest build from the summer — Daily Postcard Agent, a fully serverless AWS agent that wakes up on its own every morning, checks the real weather, writes a short poem, paints a matching postcard illustration, and publishes both to a public gallery with zero manual steps. Rather than start something new, I used the showcase weekend to take it further: close a gap I'd knowingly left open since the original build, and make the thing it's actually known for — an evolving style — real instead of aspirational.
The problem it solves hasn't changed: I like the idea of a daily creative ritual but never keep one up manually, so an agent that just does it, on a schedule, and leaves the result somewhere I can check later, removes the friction entirely. Each morning the agent pulls current weather, reads its own history of the last five days, asks Amazon Bedrock (Nova Lite) to write a title, poem, and image prompt, asks Bedrock again (Stability's Stable Image Core) to paint the postcard, and publishes the result to a gallery backed by S3, DynamoDB, and CloudFront. What's new this weekend is how it decides what "evolving style" means.
How I built it
The original version had a soft spot I flagged in my own README the day I shipped it: the
style_note that fed tomorrow's prompt was written before the image existed — a one-sentence guess about what the postcard would look like, not a description of what it actually turned out to be. It worked, but it was memory of an intention, not memory of a result. Good enough for a weekend prototype; not good enough for a showcase.So the finishing touch was a second Bedrock call, added after the image is generated rather than before it: Nova Lite is multimodal, so I pass it the actual rendered PNG — not the prompt that produced it — along with the previous day's reflection, and ask it to critique what's really on the canvas: the palette, the composition, the technique, and explicitly whether today drifted from, repeated, or broke from yesterday. That critique,
style_reflection , is what gets written to DynamoDB and what future runs read back as memory, replacing the old prospective style_note as the primary signal (I kept style_note around for backward compatibility with older records).The main implementation snag was getting the multimodal request shape right — Nova's native
invoke_model schema takes an image content block alongside text in the same message, with the image base64-encoded inline, which isn't quite the same shape as the text-only calls I already had working. Once that was sorted, I wrapped the whole critique call in a try/except that falls back to the original style_note on any failure, because a broken second call should degrade gracefully, not break the daily run. I bumped the Lambda timeout from 90s to 120s to give three sequential Bedrock calls room to breathe, redeployed the CloudFormation stack, and invoked it manually to confirm — the first post-deploy postcard's stored reflection genuinely describes its own palette and composition and gives a concrete instruction for tomorrow, not a rephrased prompt.AWS services used and architecture
- Amazon Bedrock —
amazon.nova-lite-v1:0for poem/title/prompt generation and, new this weekend, a second multimodal call on the rendered image for style critique;stability.stable-image-core-v1:1for the postcard image itself.
- AWS Lambda — the orchestrator; one function, three Bedrock calls, writes to S3 and DynamoDB.
- Amazon EventBridge Scheduler — daily cron trigger, no polling, no always-on compute.
- Amazon S3 — hosts the static gallery and postcard images, origin behind CloudFront.
- Amazon CloudFront — HTTPS in front of the S3 site, with separate cache behaviors for the daily
index.html(always revalidate) and immutable postcard images (cache aggressively).
- Amazon DynamoDB — the agent's memory: one record per day, now including the grounded
style_reflectionthe evolution loop actually reads from.
- AWS IAM / CloudFormation — least-privilege execution role, one template for the whole stack.
EventBridge Scheduler (daily cron) | v AWS Lambda ----> Open-Meteo (weather, no auth) |----> Bedrock: amazon.nova-lite-v1:0 (poem, title, image prompt) |----> Bedrock: stability.stable-image-core-v1:1 (postcard image) |----> Bedrock: amazon.nova-lite-v1:0 (multimodal) (critiques the rendered image -> style_reflection) |----> DynamoDB (history: date, theme, style_reflection, poem, image_key) |----> S3 (postcard image + regenerated index.html) ----> CloudFront (HTTPS, public URL)What I learned across the summer
Looking back, this one project ended up retracing the whole summer's arc on its own. It started as a Wishlist-shaped idea — "a daily creative ritual I'd never actually keep up manually." It became a creative app once poem and image generation were wired together. It became a real always-on agent once EventBridge and a DynamoDB-backed feedback loop replaced any notion of a human pressing a button. And this showcase weekend was the step I didn't have language for until now: the difference between an agent that claims to evolve and one that actually grounds its next decision in what it really produced, not what it meant to produce.
The broader lesson, consistent across every weekend this summer, is that the interesting engineering is rarely in the happy path — it's in the gap between intention and result. Back in the original build, that gap was Nova Canvas being unavailable in my account's Bedrock catalog, forcing a switch to Stability mid-weekend. This weekend it was the gap between a prompt describing an image and the image a diffusion model actually renders, which no amount of clever prompting can fully close — only looking at the actual output can. Building the second Bedrock call to look at its own work, rather than trust its own plan, is the single most "agentic" thing this project does, and it's the piece I'm proudest of shipping in the finale.
Try it / repo
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article