AWS Builder Center
Weekend Creative Agent Challenge: Wanderpost — A Postcard From Somewhere You've Never Been

Weekend Creative Agent Challenge: Wanderpost — A Postcard From Somewhere You've Never Been

An autonomous AI agent that generates a daily postcard from a new destination — using real weather, Bedrock Claude for evolving prose, and Nova Canvas for watercolor art — all serverless at \$0.04/day.

🌍 Vision — What It Does

What if every morning, before you even wake up, an AI traveler had already visited a new city, felt the weather on its skin, and scribbled you a postcard about it?
Wanderpost is a fully autonomous creative agent that generates a daily postcard from a place you've never been. It doesn't wait for you to ask. It doesn't need a prompt. Every morning at 6 AM, while you're still asleep, it "travels" to a new destination on a continuous slow journey across the world, checks that city's real weather, writes a short traveler's note in a voice that evolves over time, and generates a watercolor-style illustration of the scene. By the time you open your inbox or gallery, the postcard is just there.

The Problem It Solves

Most AI interactions are reactive — you ask, it answers, the conversation dies. Wanderpost flips this entirely. It's a proactive creative companion that produces something beautiful on its own schedule. It solves the "blank canvas" problem: you never have to think of a prompt, choose a destination, or specify a mood. The agent handles all of that autonomously, drawing from real-world data and its own evolving memory.
But it goes deeper than a simple daily image generator:
  • It tells a continuous story. The traveler doesn't teleport randomly — it follows curated routes (Mediterranean Drift, Silk Road Echo, Andean Spine), moving one stop per day. A week of postcards reads like an actual trip unfolding.
  • Weather shapes everything. Rain in Barcelona means introspective writing and slate-teal palettes. Snow in Tromsø means magical stillness and ice-blue art. The real weather at the destination determines both the writing's mood and the image's lighting.
  • The voice genuinely evolves. After each postcard, the agent critiques its own writing and stores a style note. Over weeks, the traveler's prose matures — it develops preferences, avoids past tics, and builds a real literary identity.
  • Friends can participate. A simple API lets anyone suggest the next stop (a place + a mood). These become "detours" in the narrative — acknowledged in the story, then the journey resumes.



🛠️ How I Built It

Process & Key Decisions

I started with a core question: How do you make an AI agent feel like it has a life of its own? The answer turned out to be three things: continuity (a slow journey, not random hops), memory (style notes that accumulate), and real-world grounding (actual weather data, not invented conditions).
Decision 1: Slow Journey Over Random Destinations
My first instinct was to pick random cities each day. But testing that felt chaotic — postcards from Tokyo, then rural Argentina, then Helsinki had no narrative thread. Instead, I designed 6 curated routes with 10-12 waypoints each. The traveler completes one route, then seamlessly begins the next. This creates the feeling of an actual trip — you can follow along on a map.
Decision 2: The Self-Critique Memory Loop
This is the heart of the project. After generating each postcard, the agent runs a second Bedrock call where it critiques its own writing: "This one leaned too melancholic — try warmer imagery tomorrow" or "The short sentence at the end worked well — keep using rhythm breaks." These notes are stored in DynamoDB and fed back into the next day's prompt. Over 2-3 weeks, the voice genuinely drifts — it's not the same every day.
Decision 3: A Recurring Companion Object
To build narrative continuity beyond just locations, the traveler carries a persistent detail — a battered leather notebook, or a brass compass that points slightly wrong. It's mentioned naturally when it fits, never forced. This small recurring element makes the postcards feel like they're written by a character, not a model.

Challenges

  • Prompt engineering for consistency: Getting Claude to write 80-150 word postcards that are consistently good (not generic, not overwrought) took multiple iterations. The trick was constraining to "one hyper-specific sensory detail per card."
  • Weather-to-art pipeline: Mapping weather conditions to artistic palettes required hand-crafting mood maps (rain → "cool silvery light, slate, deep teal, wet black"). Without this, the images felt disconnected from the text.
  • Single-table DynamoDB design: Fitting journal entries, style notes, and journey state into one table using pk/sk patterns kept the infrastructure minimal while supporting all query patterns.

☁️ AWS Services & Architecture

Architecture Overview

EventBridge Scheduler (daily, 6:00 AM IST)
│
▼
Lambda Function (Python 3.12, ARM64)
│
├─→ OpenWeatherMap API ─── real weather for destination
├─→ DynamoDB (read) ────── recent journal + style memory
├─→ Bedrock Claude 3 Haiku ── write postcard + self-critique
├─→ Bedrock Nova Canvas ──── generate watercolor scene
├─→ DynamoDB (write) ───── save entry + advance journey
├─→ S3 ─────────────────── store image + update gallery
└─→ SES ────────────────── email the postcard
API Gateway + Lambda
└─→ DynamoDB ───── community request queue

Services Used

Service
Role
Amazon EventBridge Scheduler
Triggers the Lambda once daily at 6 AM IST
AWS Lambda
Orchestrates the entire pipeline (pick destination → weather → write → generate → store → email)
Amazon Bedrock (Claude 3 Haiku)
Writes the postcard text and self-critique notes with context from journal memory
Amazon Bedrock (Nova Canvas)
Generates watercolor-style postcard images from text prompts combining scene + weather + mood
Amazon DynamoDB
Single-table design storing journal entries, style evolution notes, journey state, and community requests
Amazon S3
Hosts postcard images + a static gallery website (S3 Website Hosting)
Amazon SES
Sends the daily postcard email with embedded image
Amazon API Gateway
REST API for friends to submit destination requests
AWS SAM
Infrastructure-as-code for the entire stack

Cost: $0.04/day ($1.20/month)

The biggest cost is Nova Canvas image generation ($0.04/image). Everything else is essentially free tier.

📚 What I Learned

1. Memory Makes Agents Feel Alive

The single most impactful feature was the style-memory loop. Without it, the postcards reset daily — same tone, same patterns. With it, there's genuine drift over time. The traveler's voice matures. Day 1 postcards read differently from Day 20. This taught me that persistent memory is what separates an agent from a chatbot.

2. Constraints Breed Creativity (For AI Too)

Giving the model a 80-150 word limit, requiring exactly one sensory detail, and forcing weather to shape the mood — these constraints made the output dramatically better. Unconstrained, the model writes generic travel blog prose. Constrained, it writes vivid, punchy postcards.

3. Real-World Data as Creative Input

Using actual weather data (not hallucinated weather) grounds the fiction in reality. There's something delightful about knowing that when the postcard says "rain streaking the café windows in Lisbon," it really was raining in Lisbon that morning. It makes the fiction feel half-true.

4. Serverless is Perfect for Creative Agents

A daily creative task is the ideal serverless use case: run for 30 seconds, produce output, disappear. No idle servers, no infrastructure to maintain. The entire stack costs less than a coffee per month.

5. Single-Table DynamoDB is Underrated

Fitting three different data types (journal, style notes, state) into one table with a pk/sk pattern eliminated the need for multiple tables and simplified the SAM template considerably. For small-scale projects, single-table design is elegant and cheap.

🔗 Repository

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