AWS Builder Center

Weekend Creative Agent Challenge: Verse

An always-on AWS agent that checks the weather each morning and writes an original poem for the day — no prompts, no button, just there when you check

Vision and what it does

Last week's challenge was about building a creative app - something you open and use. This week's challenge asked for the opposite: a tool you never have to open. That constraint changed how I thought about the problem entirely. A creative app you interact with is judged by how good the interaction feels. A creative agent is judged by whether it's still doing its job when nobody's watching.
I built Verse - an agent that writes an original poem every morning, themed to that day's actual weather, and has it waiting for you before you've opened your laptop. No prompt, no button, no session. Every morning it checks the sky over Bengaluru, picks a poetic form for the week (the form rotates - haiku one week, sonnet the next, then free verse, villanelle, ode -so the "voice" visibly shifts over time instead of staying static), and asks a language model to write something that captures the specific feeling of that weather without just describing it. The poem lands in a small gallery site that reads like a weather almanac: each day is its own entry, stamped with the date, the temperature, and the poetic form used.
The problem it solves is small and personal rather than dramatic: most "AI creativity" tools require you to show up and ask. Verse flips that -it shows up for you.

How I built it

I designed this around a strict separation: one Lambda writes, one Lambda reads, and nothing else touches the model. The generator Lambda is the actual agent - on trigger, it pulls live weather data from a free API, builds a prompt around the day's conditions and the week's poetic form, and calls the model. It writes the result straight to DynamoDB keyed by date, so every day gets exactly one entry, no duplicates, no manual cleanup. The reader Lambda is dumb on purpose — it only serves what's already stored, so there's no path from the public-facing API back to an expensive model call.
The key decision was making the "self-improving style" requirement honest rather than gimmicky. I didn't want to claim real learning was happening when it wasn't. Rotating the poetic form by week is a simple, legible way to satisfy "an agent that improves its style over time" without overselling what's actually a scheduling rule, not a feedback loop.
The real challenge this weekend was infrastructure, not code. Getting IAM permissions right for a fresh AWS account (my first deploy failed on a  CreateChangeSet  AccessDenied error until I attached broader permissions to my deploy user), and then hitting AWS's automatic new-account verification hold on Bedrock calls -a security measure that blocks certain API calls for up to two hours on new accounts, unrelated to anything in my code or config. Both were reminders that "the agent works" and "the agent is allowed to run" are two separate problems on a brand-new account.

AWS services used and architecture

  • EventBridge Scheduler - fires the generator once a day, no human involved
  • AWS Lambda (Node.js 20) -two functions:  generator  (weather → prompt → Bedrock → DynamoDB) and  reader  (DynamoDB → API)
  • Amazon Bedrock (Nova Lite) -generates the poem itself
  • Amazon DynamoDB -one item per day, poem + weather metadata
  • API Gateway (HTTP API) - exposes  GET /poems  to the frontend
  • Amazon S3 - hosts the static gallery frontend as a public website
The flow:  EventBridge → Generator Lambda → Bedrock + DynamoDB , and separately,  S3 frontend → API Gateway → Reader Lambda → DynamoDB . The write path and the read path never touch the same Lambda, which kept the IAM policies for each function narrow and easy to reason about.

What I learned

Bedrock's  InvokeModel  request/response shape for Amazon Nova models is different from what I expected coming from other model APIs - worth reading the actual model card rather than assuming a familiar shape. I also came away with a better instinct for infrastructure-as-code:  sam build  +  sam deploy --guided  took a template file to a fully wired stack (DynamoDB, two Lambdas, API Gateway, EventBridge, S3, IAM roles) in one command, which made the IAM permission errors easier to diagnose since everything else was reproducible from one source of truth.

Try it / source

  • Live app: coming online as soon as Bedrock finishes account verification — link to follow
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