Banana Lab - Automated AI Prompt Pack Store

📋 Project Overview & Problem Statement

Challenge: Creating and selling digital products consistently requires weekly ideation, writing, formatting, listing, and promotion — a full-time workload for a solo operator with no team and no budget for contractors.

Solution: Banana Lab is a 3-agent Python automation system that runs every Monday to research a product idea, generate a complete 50-prompt PDF pack, update the product website, and email a weekly revenue report — with the owner only needing to convert a Markdown file to PDF and upload it to Gumroad.

Key Benefits

Zero blank-page problem: Claude picks a new niche automatically, skipping anything already in the catalogue
Production-quality output: Each pack is 50 prompts across 5 categories, written specifically for a business niche with [PLACEHOLDER] variables
Self-updating website: Site Agent rebuilds the homepage and creates a product page after every new pack
Weekly visibility: Reporter Agent emails a full P&L summary — revenue, Gumroad fees, API costs, net profit in MYR
Low running cost: ~RM 6–12/month in Claude API costs; break-even at 1–2 sales per month

🖥️ Application Features

Builder Agent

Calls Claude 3 times per cycle: once to pick a unique niche idea, once to define 5 prompt categories, then twice to generate 50 prompts in batches. Saves a clean Markdown file and ready-to-paste Gumroad listing copy.

Site Agent

Reads state.json and generates an individual product landing page for the newest pack. Then rebuilds the full homepage grid, showing all live products in reverse chronological order.

Reporter Agent

Every Sunday it reads state.json and sends an email report: products live, pending uploads, total sales, gross revenue, Gumroad fee, estimated API cost, and net profit in MYR. Flags underperformers (3+ weeks live, 0 sales).

Manual Gumroad Flow

Owner converts the Markdown file to PDF, uploads to Gumroad manually, then runs python main.py --register SLUG URL to attach the Gumroad link and mark the product live on the website.

🤖 AI Integration & Intelligence

Niche Research (Claude API)

Sends existing product titles to Claude and asks it to pick one high-demand niche not yet in the catalogue. Returns structured JSON: title, slug, description, niche, price, tags, and what's-inside bullets.

Batched Prompt Generation

Generates 50 prompts in two API calls (30 + 20) to stay within Claude's output token limits. Each prompt is niche-specific, full-length, and uses [PLACEHOLDER IN CAPS] variables for the owner's business details.

Structured JSON Output

All Claude responses are requested as raw JSON (no markdown fences). The agent strips any accidental fence wrappers and parses the response, with retry guidance built into the user guide for the rare malformed-JSON case.

Model: claude-sonnet-4-6

Uses Anthropic's latest Sonnet model for all 3 API calls per product cycle. The model is specified per call so it can be upgraded independently for each agent without changing shared config.

🛠️ Technical Architecture & Implementation

Agent Layer

BuilderAgent SiteAgent ReporterAgent StateManager

Backend & Runtime

Python 3 anthropic SDK schedule library Gmail SMTP (SSL) JSON state file

Storefront

Static HTML/CSS GitHub Pages Gumroad (payments)

System Architecture

Single state file: state.json tracks all products, sales counts, week number, and agent run timestamps — no database required
Scheduler: Python schedule library triggers Builder+Site every Monday at 09:00 and Reporter every Sunday at 18:00
Batched generation: 50 prompts split into two API calls (categories 0–2 = 30 prompts, categories 3–4 = 20 prompts) to avoid output token truncation
Manual upload gate: Products enter pending_upload status until --register is called, preventing broken "Get It Now" buttons on the live site
Self-contained website: SiteAgent writes pure HTML/CSS with no JavaScript framework — no build step, no CDN dependency

📖 Development Setup & Installation Guide

Prerequisites

Python 3.8+ with pip
Claude API key from console.anthropic.com
Gmail App Password for weekly email reports

Quick Start

# Install dependencies
pip install -r requirements.txt

# Fill in config.json with your API key and Gmail credentials

# Test run — generates first product and updates website
python main.py --builder

# Register a product after uploading to Gumroad
python main.py --register chatgpt-prompts-cafe-restaurant-owners https://yourname.gumroad.com/l/abc123

# Send a test email report
python main.py --reporter

# Start the weekly scheduler (runs indefinitely)
python main.py
            

Environment Configuration (config.json)

{
  "claude_api_key": "sk-ant-YOUR_KEY_HERE",
  "email_sender": "you@gmail.com",
  "email_app_password": "xxxx xxxx xxxx xxxx",
  "email_recipient": "you@gmail.com",
  "website_url": "",
  "state_file": "state.json",
  "products_dir": "products",
  "website_dir": "website"
}
            

Available Commands

python main.py — Start scheduler (runs indefinitely)
python main.py --builder — Run Builder + Site once
python main.py --reporter — Send one email report
python main.py --all — Run all agents once
python main.py --register SLUG URL — Register a product with its Gumroad URL

📊 Key Metrics

50

Prompts Per Pack

3

Agents Running

3

Claude API Calls Per Product

RM 6–12

Monthly API Cost

Business Value

Content at scale: 50 niche-specific, tested prompts generated in under 2 minutes of API time
Passive income pipeline: Once a product is registered, it earns indefinitely with no ongoing effort
Financial clarity: Weekly report calculates gross, Gumroad fee, API cost, and net profit automatically
Expandable architecture: Twitter promoter and Gumroad API hooks are stubbed in utils/ for future activation
Low break-even: 1–2 sales per month at $9 covers all running costs

🍌 Banana Lab