
Weekend Creative Agent Challenge: Daily Postcard Agent
An always-on AWS agent that checks the weather each morning, writes a poem, and paints a matching postcard with Amazon Bedrock — publishing the result to a public gallery before you've even opened your laptop. Built serverless on Lambda, EventBridge Scheduler, S3, and DynamoDB, with a feedback loop that nudges its own visual style forward day over day.
Weekend Creative Agent Challenge: Daily Postcard Agent
Vision and what it does
Most "creative AI" projects require you to show up: open an app, type a prompt, wait, download. The Weekend Creative Agent Challenge asked for the opposite — something that works while you're not looking, and has a result ready when you get back. I built Daily Postcard Agent, a fully serverless AWS agent that wakes up on its own every morning, checks the real weather, writes a short poem about the day, paints a matching postcard illustration, and publishes both to a public gallery — with zero manual steps after deployment.
The problem it solves is small but real: I like the idea of a daily creative ritual (a poem, a piece of art, a little moment of "what did today look like?") but I never actually keep up one manually. An agent that just does it, on a schedule, and leaves the result somewhere I can check later, removes the friction entirely. It also directly answers the challenge's own idea list — it's a daily-art agent, it themes its output to the weather, and it's designed to evolve its own style over time rather than repeat itself.
Each morning, the agent:
- Pulls current weather conditions for a fixed location.
- Reads its own history of the last five days — what themes and styles it already used.
- Asks Amazon Bedrock (Nova Lite) to write a title, a short poem, an image-generation prompt, and a one-sentence "style note" describing how today should evolve from what came before.
- Asks Bedrock again (Stability's Stable Image Core model) to paint the postcard from that prompt.
- Publishes the image and an updated gallery page to S3, and logs the day's record to DynamoDB so tomorrow's run has something to build on.
The gallery is a single static page, newest postcard first, so you can scroll back through the agent's whole run and watch its style drift over time.
How I built it
I started from the architecture, not the code: an always-on agent, by definition, needs a trigger that doesn't depend on anyone opening anything, so Amazon EventBridge Scheduler driving an AWS Lambda function was the obvious backbone — push-based, scales to zero, and free-tier friendly.
The more interesting decision was how to make "improves its style over time" real instead of cosmetic. It would have been easy to fake it with a static prompt, but instead every run reads the last five DynamoDB records and feeds their
style_note and theme fields back into the text-generation prompt, explicitly instructing the model to push the style in a new direction while keeping a recognizable throughline. It's a genuine (if simple) feedback loop: today's output shapes tomorrow's instructions.The biggest challenge wasn't code, it was model access. I'd planned to use Amazon Nova Canvas for image generation, since it's Amazon's own text-to-image Nova model and the most natural fit for the challenge's "Nova models" prompt. But once I actually tried invoking it, Bedrock returned an access error: Nova Canvas is marked
LEGACY in this account's catalog, and accounts without 30 days of prior usage are blocked from on-demand access. Rather than burn a chunk of the weekend requesting an exception, I checked the Bedrock catalog across regions and found stability.stable-image-core-v1:1 — a fully active text-to-image model — available in us-west-2 . I moved the whole stack there, kept Nova Lite for all the text generation (title, poem, prompt-writing, style notes), and used Stability for images. It's a good example of designing around what a real AWS account actually has access to, rather than the ideal case — the kind of thing you only discover by actually invoking the model, not by reading the docs.Everything else came together quickly once that was settled: a single Lambda function (Python, boto3, no extra dependencies — even the weather lookup uses the standard library's
urllib against Open-Meteo's free, keyless API) and a CloudFormation template defining the S3 bucket, DynamoDB table, IAM roles, Lambda function, and EventBridge schedule as one deployable stack.AWS services used and architecture
- Amazon Bedrock —
amazon.nova-lite-v1:0for poem/title/prompt/style-note generation,stability.stable-image-core-v1:1for the postcard image.
- AWS Lambda — the orchestrator; one function, triggered daily, calls Bedrock twice, writes to S3 and DynamoDB.
- Amazon EventBridge Scheduler — daily cron trigger (
cron(0 12 * * ? *)), no polling, no always-on compute.
- Amazon S3 — hosts the public static gallery site and the generated postcard images.
- Amazon DynamoDB — stores each day's record (theme, style note, poem, image key) as the memory the style-evolution loop reads from.
- AWS IAM / CloudFormation — least-privilege execution role and a single template deploying the whole stack reproducibly.
EventBridge Scheduler (daily cron) | v AWS Lambda ----> Open-Meteo (weather, no auth) | |----> Bedrock: amazon.nova-lite-v1:0 (poem, title, image prompt, style note) |----> Bedrock: stability.stable-image-core-v1:1 (postcard image) |----> DynamoDB (history: date, theme, style_note, poem, image_key) |----> S3 (postcard image + regenerated index.html gallery)What I learned
Model catalog access is a real design constraint, not a footnote — the model you planned around on paper (Nova Canvas, in my case) may simply not be available to a given account, and the fix is to actually probe the account's Bedrock catalog and region availability before committing to an architecture, not after. I also came away with a much better appreciation for how little infrastructure an "always-on agent" actually needs: no queue, no orchestration framework, no persistent compute — a scheduled Lambda and a DynamoDB table are enough to build something with real memory and a real feedback loop. And treating the history table as the agent's "memory of its own past work" rather than just a log turned out to be the difference between a one-shot generator and something that actually behaves like it's evolving.
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