AWS Builder Center
Weekend Creative Agent Challenge: Drift — A Daily Field Note That Learns in Public

Weekend Creative Agent Challenge: Drift — A Daily Field Note That Learns in Public

Drift is an autonomous daily creative agent powered by Amazon Bedrock and Nova. Every morning, it transforms local weather, the day, and reader reactions into a fresh micro-magazine featuring an original poem, a four-color story, and a creative prompt. It remembers audience preferences over time and clearly explains how those signals influenced each new edition.

Vision: creativity without the empty prompt
Most creative AI tools begin with an empty box. They wait for me to arrive, think of a good prompt, and ask for something. I wanted to reverse that relationship. Drift is an autonomous daily micro-magazine that creates a small, finished creative object before I wake up: a short poem, a four-color story, and one offline invitation to notice or make something.
Every issue is shaped by three kinds of context: the current local weather, the rhythm of the day, and lightweight reactions to recent issues. The reader never has to become a prompt engineer. They can simply choose “keep this direction,” “surprise me more,” “make it calmer,” or “push it bolder.” Drift carries those signals into tomorrow’s editorial brief.
The problem Drift solves is not a lack of AI-generated content. It is the friction and sameness of asking for it. Instead of producing another disposable response, Drift creates a daily ritual with continuity. Weather becomes creative material rather than a forecast, and yesterday’s reactions become compact editorial memory rather than an endless chat history.
The adaptation is deliberately visible. Each issue includes an “adaptation receipt” that explains which signals were used and how the style shifted. It does not expose private model reasoning. It gives the reader a short editorial explanation such as “quieter cadence, one vivid contrast, and more concrete imagery.” This makes the autonomous loop understandable instead of hiding it behind a sparkle icon.

How I built it

I started by defining one strict daily artifact. A successful Drift issue must contain a headline, subtitle, four-to-eight-line poem, exactly four named colors with valid hex values, a creative prompt, a weather note, and a short adaptation note. That contract became a structured prompt for Amazon Nova 2 Lite through the Amazon Bedrock Converse API.
At 5:30 each morning, Amazon EventBridge Scheduler invokes the Drift AWS Lambda function. The function first checks Amazon DynamoDB for an issue with today’s date. If one already exists, it exits, making scheduled retries safe and preventing duplicate model calls. Otherwise, it reads the seven most recent issues, aggregates their reaction counts, and fetches current conditions for the configured location. A temporary weather failure does not stop the daily run; the brief records that the signal was unavailable and continues.
Lambda then converts those inputs into a compact editorial brief. For example, more “calmer” reactions than “bolder” reactions creates guidance for a gentler cadence and more visual space. “Surprise me more” tells the model to avoid the obvious weather metaphor and introduce a less predictable conceptual turn. Recent titles are included as a do-not-repeat list.
Nova returns structured JSON, but creative model output is never written directly to storage. The Lambda function extracts and parses the JSON, checks the poem length, validates every  #RRGGBB  color, limits field sizes, and rejects incomplete output. Only a validated issue is written to DynamoDB, using a conditional write so there can be only one issue for a date.
The web interface is intentionally editorial rather than dashboard-like. It uses oversized typography, a weather dial, a daily palette, and a paper-inspired layout. The newest issue leads into the poem, color story, creative prompt, adaptation receipt, reaction controls, and archive. The public experience is hosted at the link below and reads generated issues through the Drift API. If the API is unavailable, it can still display a demonstration issue rather than becoming a blank page.

AWS services and architecture

Drift uses the following AWS services:
  • Amazon EventBridge Scheduler provides the timezone-aware daily trigger and retry policy.
  • AWS Lambda gathers context, creates the editorial brief, invokes the model, validates the result, and serves API requests.
  • Amazon Bedrock with Amazon Nova 2 Lite generates the structured creative issue.
  • Amazon DynamoDB stores immutable daily issues and compact reaction memory using on-demand capacity.
  • Amazon API Gateway HTTP API exposes read endpoints and a feedback endpoint to the website.
  • AWS SAM and AWS CloudFormation define the backend as reproducible infrastructure as code.
The architecture flow is:
 EventBridge Scheduler → Lambda → recent DynamoDB memory + current weather → Amazon Bedrock/Nova → validation → conditional DynamoDB write 
For readers, the flow is:
 Drift website → API Gateway → Lambda → DynamoDB 
Feedback travels through the same API and atomically updates the relevant issue. That signal becomes part of the next scheduled brief. The generation route is deliberately private, so a public visitor cannot trigger unexpected Bedrock usage. The system performs only one planned model request per day, while normal visitors read already-generated work. This keeps the architecture serverless, responsive, and cost-aware.

Key decisions and challenges

The hardest design question was what “improves its style over time” should mean. Letting every click rewrite a long system prompt would be unpredictable and difficult to demonstrate. Instead, I chose four bounded reactions and made their effects legible. This is not model training; it is evolving editorial memory. That distinction made the behavior safer, cheaper, and easier to explain.
Another challenge was balancing creativity with reliability. The output should feel authored and surprising, but the application still requires dependable data. Explicit output constraints plus server-side validation created a clean boundary between those goals. Conditional writes handle duplicate invocations, and the weather fallback prevents one external context source from breaking the daily ritual.
I also spent time making the visual identity support the idea. Drift should feel like a collectible field note, not an administration panel. The result is a responsive, magazine-like cover whose compass-inspired weather dial turns live conditions into part of the artwork.

What I learned

I learned that an always-on agent becomes more compelling when it has a ritual, not just a trigger. “A poem, palette, and prompt before breakfast” is a promise the reader can understand. I also learned that memory does not need to be huge. A small set of intentional signals can create a strong feeling of continuity when the interface shows what changed.
I learned to treat model output as untrusted application input. Structured prompting helps, but parsing, validation, length limits, fallbacks, idempotency, and controlled permissions are what turn a creative demo into a dependable agent.
Most of all, I learned that autonomy should be observable. A scheduled Lambda proves the process runs without me, while the adaptation receipt proves that yesterday actually matters to tomorrow.
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