AWS Builder Center

Weekend Creative Agent Challenge: Cloudy Agent

Self-executing meteorological poetry agent.

Computer Science estudent
Vision: an agent that creates while you sleep
The best tool is the one you never have to open — that line from the challenge brief stuck with me while I was scoping this project. Most "creative AI apps" still require a human to open a tab, type a prompt, and wait for output. That's not really autonomy, it's just a nicer chat window.
Cloudy Agent flips that. It's a fully autonomous creative agent that wakes up once a day, checks the weather, writes an original poem or piece of descriptive art shaped by that day's conditions, and publishes it to a public site — all without me touching anything. I go to sleep, and by morning there's something new waiting. No prompt to type, no button to press. The problem it solves isn't "how do I generate content faster," it's "how do I remove myself from the loop entirely" while still getting something worth reading.
On top of that, the agent doesn't just repeat itself. It looks back at its own three most recent generations and uses them as style references, so its voice can drift and evolve over time instead of producing the same flavor of poem every day.

How I built it

I started from a spec-driven workflow in Kiro rather than freehanding code. Instead of handing Kiro pre-made templates (which burns credits adapting them to the environment), I wrote a single detailed context prompt and had Kiro generate its own steering files,  requirements.md ,  design.md , and  tasks.md  from scratch. That upfront investment paid off — every later change (adding a cost guard, fixing a weather API choice, tightening secret handling) was a targeted edit to a living spec instead of a scattered conversation.
Key decisions along the way:
  • Direct Anthropic API instead of Bedrock. On a previous hackathon, a Bedrock account-access restriction killed a project at the final deployment step. This time, since Bedrock model access wasn't yet approved on my account, I went straight to  api.anthropic.com/v1/messages  from inside the Lambda. Same creative engine, zero dependency on an approval process I couldn't control.
  • Open-Meteo over a keyed weather API. No signup, no API key, generous free-tier limits — one less secret to manage during a weekend timeline.
  • AWS Secrets Manager over environment variables. The Claude API key is fetched at Lambda cold start via  GetSecretValue , never baked into the deployment package or exposed as a plaintext env var. I also made sure the IAM user driving Kiro's deploys never needed permission to read that secret — only to create the surrounding infrastructure.
  • A single orchestrator Lambda with an early-exit pipeline (weather → context → style memory → cost guard → generate → store → publish), rather than a multi-agent setup. Given the weekend scope, one well-structured function beat the coordination overhead of several.
The biggest challenge wasn't the AI part — it was the boring infrastructure edges: DynamoDB billing mode defaults, S3 buckets that silently 403 because Block Public Access is on by default, and making sure a failed run never leaves the public site half-published. Property-based tests (via  hypothesis ) ended up catching more of these edge cases than example-based tests would have, especially around the cost guard and the "never log the API key" guarantee.

AWS services and architecture

  • Amazon EventBridge Scheduler — daily cron trigger, concurrency capped at 1
  • AWS Lambda (Python 3.12) — the orchestrator; calls Open-Meteo, the Claude API, DynamoDB, and S3 in sequence
  • AWS Secrets Manager — holds the Anthropic API key, fetched at cold start
  • Amazon DynamoDB — stores every generation plus token usage, with a GSI for recent-style lookups and monthly cost aggregation
  • Amazon S3 — static website hosting for the daily output page
  • AWS SAM — single  template.yaml , one  sam deploy , full stack reproducible
EventBridge → Lambda → Secrets Manager (key)
→ Open-Meteo (weather)
→ Claude API (generation)
→ DynamoDB (memory + cost tracking)
→ S3 (publish)


What I learned

Spec-driven development with an agentic IDE is dramatically more efficient once you invest in a solid context prompt upfront — the cost isn't in writing the spec, it's in not writing one and re-explaining context every session. I also came away with a much sharper instinct for where "free tier" projects quietly leak cost or break (GSI design, DynamoDB billing mode, S3 public access), and why a Cost Guard isn't optional when an LLM call runs unattended on a schedule you don't personally watch.

Try it

Live site:  http://agent-aws-staticsitebucket-ckleey66n3ww.s3-website-us-east-1.amazonaws.com/  Repo: (link coming soon — repo is being finalized)
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