
Weekend Creative Agent Challenge: Prompt Garden
Prompt Garden is an always-on creative agent that turns community prompt seeds into a fresh daily creative challenge and AI-generated inspiration image. AWS EventBridge triggers the workflow automatically, while Lambda orchestrates GPT-5 mini, FLUX.2 Pro, DynamoDB, S3, and a public gallery.
Weekend Creative Agent Challenge: Prompt Garden
Tag: #agents
Vision And What It Does
Creative prompts are easy to find, but a blank page is still intimidating. I wanted to make a tool that removes the “what should I make today?” decision without replacing human creativity. Prompt Garden is an always-on creative agent that prepares one fresh community-inspired challenge before anyone opens the app. When a visitor returns, a new visual and written brief is waiting for them.
Prompt Garden starts with a small garden of community-style prompt seeds, such as “a forgotten train route through the clouds,” “a tiny astronaut living inside a houseplant,” and “a robot learning to grow a garden.” Visitors can also plant their own seed in the deployed gallery. The next scheduled run prioritizes the three newest submissions, then fills any remaining slots with built-in seeds. Every day, the agent remixes its selected seeds into a single original prompt. It creates a title, a concise challenge suitable for drawing, writing, photography, or design, visual art direction, and a note explaining how the source seeds were combined. It then creates an inspiration image that matches the challenge.
The first successful autonomous output was titled Cloudline Garden Express. It combined a sky train, a tiny astronaut in a houseplant, and a learning gardening robot into a creative brief about a drifting greenhouse community above the clouds. The important point is that this output was created by an EventBridge event, not by a user pressing a button. The generated record identifies its trigger source as
aws.events .How I Built It
I chose to make the agent backend-first. A polished frontend is useful, but the heart of this challenge is reliable autonomous behavior. I began with an EventBridge scheduled rule that invokes an AWS Lambda function daily at 08:00 UTC. Lambda selects three seeds, requests structured text output from OpenAI GPT-5 mini, requests an image from Azure AI Foundry FLUX.2 Pro, saves the PNG to S3, and writes the complete daily artifact to DynamoDB.
I used a strict JSON schema for the text response. This was an important decision because the Lambda needs dependable fields rather than prose that must be guessed or parsed. The model returns
title , challenge , art_direction , and remix_note , which makes the DynamoDB record predictable and gives the image model a clean art-direction prompt.One challenge was model access. Amazon Bedrock model access was unavailable in the AWS account, even though the deployed Lambda role had the correct IAM permission. Rather than abandoning the always-on AWS architecture, I kept the scheduling, storage, orchestration, and logs on AWS while using external model APIs. Another issue was that GPT-5 mini initially exhausted a response budget before returning visible JSON. Increasing the output-token budget resolved that. I also corrected the Azure Foundry request logic to accept a full FLUX.2 provider endpoint.
Secrets never enter the repository. Lambda retrieves the OpenAI and Azure credentials at runtime from AWS Secrets Manager. The private S3 bucket also keeps generated images protected; the public gallery Lambda creates a short-lived signed image URL only when someone views the latest challenge.
AWS Services And Architecture
The application is deployed on AWS and uses several AWS services:
- Amazon EventBridge is the always-on scheduler.
- AWS Lambda runs the creative agent and serves the public gallery through a Lambda Function URL.
- Amazon DynamoDB persists the daily title, challenge, selected seeds, model metadata, image key, time, and trigger source.
- Amazon S3 stores generated PNG images privately, with a 30-day lifecycle rule.
- AWS Secrets Manager stores external API credentials securely.
- Amazon CloudWatch Logs provides execution evidence and operational debugging information.
The architecture is straightforward: EventBridge invokes the generator Lambda; Lambda calls GPT-5 mini for a structured challenge and FLUX.2 Pro for an image; the resulting record goes to DynamoDB and the image goes to S3. A second Lambda Function URL reads the newest record, signs the private image for temporary display, and returns the Prompt Garden gallery page.
What I Learned
This project reinforced that an agent is more than a model call. Scheduling, durable storage, secrets management, observability, and safe public delivery are what turn a generation script into an always-on product. I also learned the value of verifying the exact event source in stored output. Seeing
aws.events in DynamoDB and CloudWatch is much stronger evidence of autonomy than manually invoking a Lambda and claiming it is scheduled.I also learned to design for provider changes. The AWS orchestration layer remains stable while the text or image provider can change independently. That separation kept the project moving when a Bedrock account-level access limitation appeared. Prompt Garden now has a dependable AWS backbone and a clear path for future additions such as real community seed submissions, voting, themed days, and style preferences that improve over time.
Links:
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article