Forem: fracabu

I Built a Prompt Framework Explorer With AI. Then the AI Created Its Own Frameworks.

fracabu — Tue, 13 Jan 2026 13:26:49 +0000

The Recursive Loop Nobody Expected

I started Promptify as a personal lab—a way to stop drowning in 47 browser tabs every time I needed to test a prompt framework.

Then something weird happened.

I built the tool using Claude Code and GLM 4.6. And when I asked them to create experimental frameworks optimized for modern LLMs, they generated:

CALIBRO (modular, enterprise-grade)
BCM (GLM-4.6 optimized)
PRISMA (Claude-optimized, multi-path reasoning)

Now those AI-generated frameworks are inside the tool, being used to test... the AIs that created them.

Meta? Absolutely.
Useful? Surprisingly, yes.

What It Does

Promptify Vue centralizes 30+ prompt engineering frameworks in one place. But unlike other collections, it includes frameworks that were designed by AI, for AI.

👉 Try the live demo
👉 GitHub repo

Core Features

1. Browse 30+ Frameworks by Category

Organized in 7 categories from Fundamentals to Advanced Systems:

Fundamentals (7): APE, BAB, PAR, RTF
Innovative (2): CALIBRO, PRISMA ← AI-generated
Business & Professional (5): AIDA, SMART, STAR
Creative & Marketing (5): CRAFT, ROSES, SPARC
Advanced & System (7): Chain-of-Thought, Tree-of-Thoughts
Problem Solving (7): Structured methodologies

2. Multi-Provider Testing

Test across OpenAI, Google Gemini, and ZAI to see how different models interpret the same framework.

3. Built With AI

The entire codebase was developed using:

Claude Code (agentic coding CLI)
GLM 4.6 (Chinese LLM for specific optimizations)

Then I asked them to design frameworks. They came up with CALIBRO (modular, self-evaluating) and PRISMA (meta-framework combining ReAct, Tree-of-Thoughts, and enterprise patterns).

Tech Stack

Category	Technology
Frontend	Vue 3.5, TypeScript 5.6, Vite 7.1
Styling	Tailwind CSS 3.4
State	Pinia
Icons	Lucide, Heroicons
Deployment	Vercel
Development Partners	Claude Code, GLM 4.6

What I Learned

1. AI-generated frameworks are different

CALIBRO and PRISMA have a level of self-awareness that human-designed frameworks lack. They include meta-evaluation steps ("Did I answer correctly? Check X, Y, Z").

2. The best framework isn't always the most popular

After daily use, I've found BAB (Before-After-Bridge) and ROSES outperform Chain-of-Thought for creative and marketing tasks. But nobody talks about them.

3. Cross-provider consistency varies wildly

Same prompt + framework → completely different outputs across OpenAI vs Gemini vs ZAI. Testing is mandatory.

4. Personal labs evolve

What started as "stop opening 50 tabs" became a research tool that contributed back to the AI ecosystem.

The Recursive Part

Here's the kicker:

I use PRISMA (created by Claude) to generate prompts for Claude Code to build features for Promptify... which contains PRISMA.

┌─────────────────────────────────┐
│  Claude creates PRISMA          │
│         ↓                       │
│  PRISMA generates prompts       │
│         ↓                       │
│  Prompts guide Claude Code      │
│         ↓                       │
│  Claude Code builds Promptify   │
│         ↓                       │
│  Promptify contains PRISMA  ←───┘

It's frameworks all the way down.

Getting Started

git clone https://github.com/fracabu/promptify-vue.git
cd promptify-vue
npm install
npm run dev

Open http://localhost:5177

What's Next

I'm exploring:

Framework Comparison Mode: A/B test multiple frameworks side-by-side
AI Framework Selector: Input your task → get recommended framework
Benchmark Dataset: Public data for framework research

But honestly? The most interesting part is watching AI contribute to its own evolution.

Your Turn

Questions for the community:

Have you used AI to build development tools? How meta did it get?
What's your go-to prompt framework? (Popular or hidden gem?)
Should I ask Claude to design a better CALIBRO? (Is this how Skynet starts?)

Drop your thoughts below! 👇

👤 Francesco Capurso (@fracabu)
Self-taught dev | AI agents & Fastify plugins

🔗 GitHub | npm

⭐ Star the repo if you find it useful

README2Video - Transform Any GitHub README into a Professional Promo Video

fracabu — Mon, 05 Jan 2026 07:00:41 +0000

This is a submission for the DEV's Worldwide Show and Tell Challenge Presented by Mux

What I Built

README2Video is a web application that automatically transforms GitHub README files into professional promotional videos.

README.md → AI Analysis → Video Generation → Mux Streaming → Share!

Paste a GitHub URL, choose your style, and get a cinematic promo video in minutes. No editing skills required.

My Pitch Video

stream.mux.com

Meta twist: This pitch video was generated by README2Video itself, from its own README. The app pitched itself. 🎬

Video Preview (Powered by Mux)

Here's an animated preview generated automatically by Mux:

Demo

Live App: readme-to-video.onrender.com
GitHub Repo: github.com/fracabu/readme-to-video

How to Test (2 minutes setup, all free!)

The app uses BYOK (Bring Your Own Keys) - this keeps costs transparent and gives you full control.

Service	How to Get Key	Cost
Mux	mux.com - Create free account	Free (no credit card)
OpenRouter	openrouter.ai - Sign up, get API key	Free models available!
Kie.ai	kie.ai - Sign up, get API key	Free credits on signup

Free models tip: In OpenRouter, select models ending with :free like deepseek/deepseek-r1:free or google/gemini-2.0-flash-exp:free.

The Story Behind It

Every developer knows this pain: you build something amazing, but creating a promo video takes hours (or days), costs money, and requires skills you don't have.

I wanted to solve this for myself and every developer out there.

The solution? Let AI do the heavy lifting:

Analyze the README to understand what the project does
Generate a compelling script with scenes
Create stunning visuals with Sora 2
Host on Mux for instant streaming

Now any developer can go from README to promo video in under 5 minutes.

The Opportunity

There are 300+ million repositories on GitHub. Every single one could benefit from a promo video.

Indie developers launching products
Open source projects seeking contributors
Startups pitching to investors
Students showcasing portfolios

README2Video turns a $500+ video production into a $2 automated process. That's the kind of democratization that scales.

Technical Highlights

Architecture

Layer	Technology
Frontend	Next.js 14, React 18, Tailwind CSS, shadcn/ui
LLM	Multi-provider: Claude, GPT-4, Gemini, OpenRouter (free tier!)
Video AI	Kie.ai Sora 2 - cinematic text-to-video
Streaming	Mux Video API
Real-time	Server-Sent Events for live progress

Key Features

Multi-Provider LLM: Choose between Anthropic, OpenAI, Google, or free models via OpenRouter
Video Styles: Tech, Minimal, or Energetic aesthetics
Durations: 15s, 30s, or 60s videos
Real-time Progress: SSE updates showing each generation step
BYOK Model: Users bring their own API keys - no hidden costs, full transparency
Dynamic Thumbnail Slider: Interactive time picker (0-15s) to select the perfect thumbnail
Animated GIF Preview: One-click GIF generation for social sharing
MP4 Download: Direct download button for offline use

Use of Mux (Additional Prize Category)

README2Video leverages 6 distinct Mux features - all exposed in the UI for users to interact with:

1. Programmatic Upload API

Videos generated by Sora 2 are uploaded directly to Mux via the Node SDK:

import Mux from '@mux/mux-node';

const mux = new Mux({
  tokenId: process.env.MUX_TOKEN_ID,
  tokenSecret: process.env.MUX_TOKEN_SECRET,
});

// Upload generated video to Mux
const asset = await mux.video.assets.create({
  input: [{ url: generatedVideoUrl }],
  playback_policy: ['public'],
  encoding_tier: 'baseline'
});

2. Adaptive Bitrate Streaming

Mux automatically transcodes every video into multiple resolutions and bitrates. Users get smooth playback regardless of their connection:

// Mux handles all transcoding automatically
// Users stream at optimal quality for their bandwidth
const streamUrl = `https://stream.mux.com/${playbackId}.m3u8`;

3. Playback URL Generation + MP4 Download

Each video gets a unique, shareable streaming URL - plus a direct download option:

// Get the playback ID from the asset
const playbackId = asset.playback_ids[0].id;

// Generate streaming URL (HLS format)
const streamUrl = `https://stream.mux.com/${playbackId}.m3u8`;

// Generate direct MP4 for download
const mp4Url = `https://stream.mux.com/${playbackId}/medium.mp4`;

In the app: A prominent "Download MP4" button lets users save their video locally for offline use or uploading to other platforms.

4. Dynamic Thumbnails with Interactive Slider

Mux generates thumbnails at any timestamp. We built an interactive slider so users can scrub through their video and pick the perfect frame:

// Thumbnail at specific time (in seconds)
const thumbnail = `https://image.mux.com/${playbackId}/thumbnail.png?time=5`;

// Custom dimensions
const customThumb = `https://image.mux.com/${playbackId}/thumbnail.png?time=10&width=1280&height=720`;

In the app: Users drag a slider (0-15s) and see the thumbnail update in real-time. One click copies the URL for social sharing.

Example - Dynamic thumbnail from our pitch video:

5. Animated GIF Previews

Generate eye-catching GIF previews for social media and documentation:

// Animated GIF from any segment
const gifPreview = `https://image.mux.com/${playbackId}/animated.gif?start=0&end=10&width=480`;

// Customize FPS for smaller file size
const optimizedGif = `https://image.mux.com/${playbackId}/animated.gif?start=2&end=8&fps=15&width=320`;

In the app: GIF previews are lazy-loaded (click to view) to save bandwidth. Users can copy the GIF URL instantly for sharing on Twitter, Discord, or README files.

Example - Animated preview of our pitch video:

6. Embeddable Player

Users get one-click access to embed code for any platform:

<!-- Mux Player embed -->
<iframe
  src="https://player.mux.com/{playbackId}"
  width="640"
  height="360"
  frameborder="0"
  allowfullscreen>
</iframe>

<!-- Or use the stream URL directly -->
<video controls>
  <source src="https://stream.mux.com/{playbackId}.m3u8" type="application/x-mpegURL">
</video>

Why Mux Was the Right Choice

Requirement	Mux Solution
Fast integration	Clean SDK, 10 lines of code to upload
No infrastructure	Mux handles encoding, CDN, player
Auto-optimization	Adaptive streaming out of the box
Rich media features	Thumbnails, GIFs, storyboards - all automatic
Developer experience	Excellent docs, predictable API
Fair pricing	Pay only for what you use

Building README2Video with Mux took minutes instead of days. That's exactly the developer experience I wanted to pass on to my users.

Challenges I Faced

1. Video Concatenation

Generating multi-scene videos (30s, 60s) required stitching multiple Sora 2 clips. Solution: FFmpeg with crossfade transitions on the server.

2. Long Generation Times

60-second videos (4 scenes) can take 5+ minutes. Solution: Server-Sent Events for real-time progress updates so users know exactly what's happening.

3. Free Tier Hosting Timeouts

Serverless platforms timeout during long generations. Solution: Deployed on Render as a persistent Node.js process.

What's Next

[x] ~~Dynamic Thumbnail Slider~~ - Done! Interactive time picker in result view
[x] ~~Animated GIF Preview~~ - Done! Lazy-loaded GIF for social sharing
[x] ~~MP4 Download~~ - Done! One-click download button
[ ] Mux Player React component for enhanced playback controls
[ ] Auto-generated captions using Mux's transcription API
[ ] Storyboard preview before final generation
[ ] Webhook notifications when video is ready
[ ] Batch processing for multiple READMEs

Built with ❤️ and AI | Powered by Mux, OpenRouter, Kie.ai & Sora 2

By submitting this project, I confirm that my video adheres to Mux's terms of service: https://www.mux.com/terms

From 0 to 500 Free Pages Scraped with Firecrawl MCP Server and Claude Code

fracabu — Fri, 02 Jan 2026 11:04:44 +0000

0 lines of code. That's what I wrote to scrape 500 pages. With Firecrawl's MCP server integrated into Claude Code, I just describe what I need in plain English, and the AI handles the rest. When I discovered this workflow combined with Firecrawl's free tier, my entire approach to web scraping changed forever.

Firecrawl's Generous Free Tier

Before diving into the technical setup, let me highlight why Firecrawl stands out. Most scraping APIs require a credit card or offer crippled free tiers. Firecrawl is different:

Plan	Price	Credits	Credit Card Required
FREE	$0	500 (one-time)	No
Hobby	$16/mo	3,000/mo	Yes
Standard	$83/mo	100,000/mo	Yes
Growth	$333/mo	500,000/mo	Yes

500 credits to start, completely free, no strings attached. That's enough to scrape an entire e-commerce category, extract data from dozens of documentation sites, or test your scraping workflow before upgrading.

The Magic Stack

Component	What It Does
Firecrawl Free Tier	500 credits to start, $0, full API access
Firecrawl MCP Server	Exposes Firecrawl API to AI assistants
Claude Code	AI that calls MCP tools autonomously

The result? You talk to Claude, Claude talks to Firecrawl, you get structured data back. No Postman, no curl commands, no boilerplate code.

Setting Up Firecrawl MCP Server

1. Get Your Free API Key

Go to firecrawl.dev and sign up. You get 500 free credits to start, no credit card required.

2. Add MCP Server to Claude Code

In your Claude Code configuration:

{
  "mcpServers": {
    "firecrawl": {
      "command": "npx",
      "args": ["-y", "firecrawl-mcp"],
      "env": {
        "FIRECRAWL_API_KEY": "fc-your-key-here"
      }
    }
  }
}

3. Start Scraping with Natural Language

Now just ask Claude:

"Scrape this URL and extract the main content"
"Map all links on this website"
"Extract product names and prices from this page"
"Crawl this documentation and summarize each section"

Claude automatically uses Firecrawl MCP tools to fulfill your request.

Why I Also Built a Visual UI

While MCP + Claude Code is incredibly powerful for ad-hoc scraping, I built Firecrawl Power App for cases where I need:

Visual feedback on what's being scraped
Persistent history of past scrapes
Quick access without opening Claude Code
Sharing results with non-technical teammates

fracabu / firecrawl-power-app

Powerful web UI for Firecrawl API + MCP Server docs. Scrape, Crawl, Map, Search, Extract & AI Agent. Built with React 18 + Vite + Claude Code.

🔥 Firecrawl Power App

▶️ Click to watch demo video

A sleek, powerful UI to harness the full potential of Firecrawl API

Features • Demo • Quick Start • Deployment • Italiano

✨ Features

Feature	Description	Plan
📄 Scrape	Extract content from any webpage (Markdown, HTML, Screenshots)	Free
🗺️ Map	Discover all URLs on a website	Free
🕷️ Crawl	Crawl entire websites with depth control	Free
🔍 Search	Search the web with advanced operators	PRO
📊 Extract	Extract structured data using AI and JSON schemas	PRO
🤖 Agent	Autonomous AI agent for complex data gathering	5/day free
🌙 Dark Mode	Beautiful dark/light theme toggle	Free
🔑 BYOK	Bring Your Own Key - use your personal Firecrawl API key	Free

Note: PRO features require a paid Firecrawl plan. Free tier includes 500 credits (one-time).

🎨 Design

Built with the official Firecrawl design system:

--primary: #FF4C00;      /*

…

View on GitHub

🚀 Try it live on Vercel

The app complements the MCP workflow. Use Claude Code for complex, one-off extractions. Use the UI for repetitive tasks and visual exploration.

App Features

The Power App provides dedicated panels for every Firecrawl operation:

Feature	Description	Plan
Scrape	Extract content from single pages	Free
Map	Discover all links on a website	Free
Crawl	Systematically index entire sites	Free
Search	Find specific content across the web	PRO
Extract	AI-powered structured data extraction	PRO
Agent	Autonomous AI agents for complex tasks	5/day free

Category	Technologies
Frontend	React, Vite, Custom Design System
Backend	Express, Vercel Serverless Functions
AI	Firecrawl API (Extraction, Agent)

My Results After 3 Months

Metric	Before	After
Time to first scrape	30+ minutes	30 seconds
Lines of code written	100+ per project	0
Monthly pages scraped	~200	480+
Cost	$0 (but hours of work)	$0 (and minutes of work)

Pro Tips

Use Claude Code for complex extractions: When you need to scrape + transform + analyze, let Claude chain the operations. Ask: "Scrape this page, extract the data, and create a CSV file."
Combine Map + targeted Scrape: First ask Claude to map a site, then selectively scrape only the pages you need. This maximizes your free credits.
Leverage JSON Schemas for precision: Use the Extract feature with a defined JSON schema to get perfectly structured, machine-readable data without parsing.

What Would You Scrape with AI + Free Credits?

The combination of MCP servers and free API tiers is opening up possibilities that didn't exist a year ago. What would YOU build? A research assistant? A content aggregator? An automated monitoring system? Share your ideas below!

👤 Francesco Capurso (@fracabu)
Self-taught dev | AI agents & Fastify plugins

🔗 GitHub | npm

⭐ Found this useful? Star the repo!

From 3 Hours to 10 Minutes: How I Built DevFlow AI to Automate Technical Writing

fracabu — Tue, 23 Dec 2025 08:00:00 +0000

3 hours. That's how long I spent crafting each technical article before DevFlow AI. After missing deadlines for my open-source projects yet again, I built an AI co-pilot for technical content creation.

The Problem

As a solo developer juggling coding and content:

Manual extraction pains: Copy-pasting READMEs killed my workflow
SEO overhead: Optimizing articles felt like guesswork
Consistency issues: Voice/tone varied wildly between rush jobs
Engagement debt: Crafting hooks took longer than writing

The Solution

DevFlow AI - transforms GitHub repos into polished technical articles:

fracabu / devflow-ai

Transform GitHub repositories into SEO-optimized technical articles. Multi-provider AI (OpenRouter + Gemini), 18+ models, Dev.to ready output. React 19 + TypeScript + Vite.

🚀 DevFlow AI

Transform GitHub Repositories into SEO-Optimized Technical Articles

▶️ Click to watch demo video

Italiano • English

🇮🇹 Italiano

Panoramica

DevFlow AI è una piattaforma che trasforma repository GitHub in articoli tecnici di alta qualità, ottimizzati per SEO. Costruito per sviluppatori che vogliono scalare il content marketing senza sacrificare tempo di coding.

✨ Caratteristiche

Funzionalità	Descrizione
🔄 GitHub Sync	Estrazione istantanea di metadata e README da qualsiasi repository
🔑 GitHub Token	Supporto token per repo private e rate limit aumentato (100 repo)
🧠 Multi-Provider AI	Supporta OpenRouter (18+ modelli) e Google Gemini
🆓 Modelli GRATUITI	Usa Gemini 2.0 Flash, DeepSeek R1, Llama 3.3 senza costi
✍️ Storytelling Hook	Articoli con intro personale e domanda engagement finale
💬 Primo Commento	Genera commento da postare dopo la pubblicazione
📅 Editorial Planner	Archivia e gestisci la pipeline di contenuti
🌍 Bilingue	Supporto completo Italiano e Inglese
📤 Dev.to Direct	Pubblica bozze

…

View on GitHub

)

How It Works

🔄 GitHub Sync Engine

Instantly fetches repository context:

// Parses markdown metadata
const extractMetadata = (readme: string) => {
  const titleMatch = readme.match(/^# (.+)/);
  return {
    title: titleMatch ? titleMatch[1] : 'Untitled',
    description: ''// Auto-extracted
  };
};

Automates repository scanning via GitHub API, supporting private repos with token auth.

🧠 Multi-AI Content Engine

Generates SEO-ready articles using optimal model per task:

Task	Model	Why
Filtering	Mistral Small 3.1	Fast topic selection
Drafting	Gemini 2.0 Flash	Balanced technical depth
Sections	DeepSeek R1	Structural coherence

Cost-efficient workflow averages $0.12/article using free tier models.

📤 Dev.to Pipeline

1-click publishing with automatic frontmatter:

---
title: "Automating Docs with AI"
published: false
tags: ["ai", "devtools", "productivity"]
---

Includes generated first comment optimized for engagement.

🛠️ Tech Stack

Category	Technologies
Frontend	React 19, TypeScript 5, Vite
Styling	Tailwind CSS (CDN), Lucide icons
AI Providers	OpenRouter, Google Gemini API
Storage	localStorage editorial calendar

My Results

Quantifying content creation transformation:

Metric	Before	After
Time per article	3+ hours	10 minutes
SEO optimizations	Manual	AI-validated
Engagement hooks	Generic	Personalized
Publication pipeline	Ad hoc	Scheduled queue

Saved 22+ hours monthly while doubling output quality ratings.

Pro Tips

Model Cost Control: Stick with free Gemini/DeepSeek unless handling complex architectures
Token Leverage: GH tokens unlock access for private repos & avoid rate limits

Engagement Question

What's YOUR biggest bottleneck in technical content creation?

👤 Francesco Capurso (@fracabu)
Self-taught dev | AI agents & Fastify plugins

🔗 GitHub | npm

⭐ Found this useful? Star the repo!

I Built a Squad of AI Agents to Chase Invoices for Me

fracabu — Sun, 21 Dec 2025 20:37:36 +0000

Chasing clients for money is the absolute worst part of freelancing and small business management. It's awkward, it consumes hours of mental energy, and quite frankly, I'd rather be coding than writing "friendly reminders" for the tenth time.

But as a self-taught dev who used to work in a warehouse, I know that if a task is repetitive and rule-based, it can be automated.

So, instead of hiring a virtual assistant, I built a team of AI agents to do the dirty work for me.

🛠️ The Problem

Before building this tool, my process was a mess. I had a spreadsheet full of invoice dates. Every Monday, I had to:

Check who hadn't paid.
Calculate how many days they were overdue.
Decide the tone of the email (Friendly? Firm? Legal threat?).
Draft the email manually to avoid sounding like a robot.
Handle the replies ("I sent it yesterday!" or "Can I pay next week?").

It wasn't just about the time; it was the context switching. Stopping deep work to play debt collector destroys productivity.

🚀 The Solution: PayMind

I built PayMind, an autonomous system where three specialized AI agents collaborate to manage the entire payment lifecycle.

I didn't just want a script that blasts generic templates. I needed intelligence. I used Next.js 16 for the dashboard, TypeScript for safety, and Claude Code to orchestrate the agents.

Crucially, I integrated OpenRouter to access 18+ models, meaning I can run the heavy lifting on free models (like Gemini 2.0 Flash) and save the premium compute (Claude Sonnet 4) for complex negotiations.

fracabu / Agent-PayMind

PayMind is an AI payment reminder system that autonomously analyzes invoices, generates personalized messages, and handles customer responses using Claude AI agents.

💰 PayMind

AI-Powered Payment Reminder System

Italiano • English

🇮🇹 Italiano

Panoramica

PayMind è un sistema intelligente di gestione automatica dei solleciti di pagamento basato su Agenti AI Claude Code. Utilizza tre agenti specializzati che lavorano in team per analizzare fatture, generare messaggi personalizzati e gestire le risposte dei clienti.

✨ Caratteristiche

Funzionalità	Descrizione
🤖 3 Agenti AI Specializzati	Team di agenti che collaborano per gestire l'intero workflow
📊 Analisi Intelligente	Identifica fatture scadute, calcola priorità e segmenta clienti
📧 Multi-Canale	Genera messaggi per Email, SMS e WhatsApp
💬 NLP Avanzato	Analizza risposte con intent recognition e sentiment analysis
🎨 Dashboard Moderna	Interfaccia Next.js per visualizzare workflow in tempo reale
🔌 Multi-Provider AI	Supporta Anthropic, OpenAI, OpenRouter e Google Gemini
🗄️ Database SQLite	Persistenza dati con Prisma ORM
🌙 Dark/Light Mode	Toggle tra tema scuro e chiaro
🌍 Bilingue	Supporto completo Italiano e Inglese
📊 Grafici Interattivi	Dashboard con

…

View on GitHub

⚙️ How It Works

The architecture relies on agent specialization. One massive prompt never works as well as three focused ones.

1. The Payment Monitor (The Analyst)

This agent scans the CSV/Database. It identifies overdue invoices, calculates priority scores (based on amount and days overdue), and segments clients. It doesn't write emails; it just serves data.

2. The Reminder Generator (The Copywriter)

This agent takes the data and context. If the client is 3 days late, it drafts a polite nudge. If they are 60 days late, it drafts a formal notice. It adapts to the preferred channel (Email, SMS, WhatsApp).

3. The Response Handler (The Negotiator)

This is the coolest part. When a client replies, this agent analyzes the intent using NLP.

Client says: "The bank transfer is scheduled for tomorrow."
Agent detects intent: PAYMENT_PROMISED
Agent action: Pauses further reminders for 48h.

All of this is visualized in a modern Next.js dashboard with a dark mode (because we're developers, obviously).

📊 My Results

Here is the breakdown of managing 50 active invoices per month manually versus using PayMind.

Metric	Manual Approach	PayMind Automation
Time Spent	~5 hours/month	~10 minutes/month
Context Switching	High (Daily interruptions)	Zero (Batch processing)
Tone Consistency	Variable (Depends on my mood)	Perfect (Always professional)
Cost	My hourly rate x 5	~$0.10 (API Costs) or FREE
Stress Level	High	Non-existent

💡 Pro Tips

If you fork this repo to build your own agent system, keep these in mind:

Model Routing is Key: You don't need GPT-4 or Claude 3.5 Sonnet to parse a CSV date. I configured PayMind to use Gemini 2.0 Flash (Free) via OpenRouter for the analysis phase. Only use the expensive models for generating the actual human-like text. This cuts costs by 90%.
Human-in-the-Loop: Always start with a "Draft Mode." PayMind generates the emails but waits for a final approval click on the dashboard before sending. AI is great, but you don't want it hallucinating a legal threat to your best client because of a CSV typo.

Conclusion

I didn't go to a bootcamp to learn this. I used curiosity and AI assistants to build an AI assistant. It's meta, but it works.

Automating the boring stuff is the best way to free up time to learn more cool stuff. Check out the code, star the repo, and stop chasing invoices manually.

👤 Francesco Capurso (@fracabu)
Self-taught dev | AI agents & Fastify plugins

🔗 GitHub | npm

⭐ Found this useful? Star the repo!

Stop paying $20/month for AI videos. Generate Sora 2 quality clips for just $0.15 each with Kie.ai API

fracabu — Wed, 17 Dec 2025 16:24:03 +0000

Hey everyone! 👋

I want to share how I went from spending $20/month on AI video tools to generating the same quality videos for $0.15 each.

If you've ever wanted to create AI-generated videos but felt priced out by OpenAI Sora or Runway, this is for you.

💸 The Problem: AI Video is Expensive

Creating promotional videos has always been costly:

Option	Cost
Freelancer/Agency	$500+ per video
OpenAI Sora	$20/month (limited)
Runway Gen-3	$0.50+ per clip
Pika Labs	$0.40+ per clip

I needed videos for my projects, but couldn't justify these costs for experimentation.

🎯 The Solution: Kie.ai API

Then I discovered Kie.ai — they provide API access to Sora 2 at a fraction of the cost:

Cost Comparison

Service	15-second video
OpenAI Sora	~$2-5
Runway Gen-3	$0.50+
Kie.ai Sora 2	$0.15

That's 100x cheaper. Same model, same quality.

🛠️ What I Built

Instead of just using the API manually, I built a complete pipeline:

📝 Describe your business
        ↓
🤖 Claude AI generates storyboard (4 scenes)
        ↓
🎬 Kie.ai Sora 2 renders each scene
        ↓
🔗 FFmpeg merges everything
        ↓
🎥 60-second promo video

Total cost for a 4-scene promo: ~$0.63

Compare that to $500+ with traditional methods.

fracabu / kie-ai-app

Automates the entire process of generating promotional videos, from AI-powered storyboard creation to video generation with Kie.ai Sora 2 and final merging.

Kie Ai App

Automates the entire process of generating promotional videos, from AI-powered s

🇬🇧 English | 🇮🇹 Italiano

🇬🇧 English

Overview

Automates the entire process of generating promotional videos, from AI-powered storyboard creation to video generation with Kie.ai Sora 2 and final merging.

Quick Start

git clone https://github.com/fracabu/kie-ai-app.git
cd kie-ai-app

🇮🇹 Italiano

Panoramica

Automates the entire process of generating promotional videos, from AI-powered storyboard creation to video generation with Kie.ai Sora 2 and final merging.

Avvio Rapido

git clone https://github.com/fracabu/kie-ai-app.git
cd kie-ai-app

Tech Stack

Python

License

MIT

View on GitHub

✨ Why Kie.ai Works

🔑 Key Benefits

Feature	Details
Same Quality	Uses actual Sora 2 model
Pay Per Use	No monthly subscription
Simple API	Just 3 endpoints
Fast	2-5 minutes per video
No Watermark	Professional output

⚙️ How the API Works

The API uses a simple async pattern:

Create Task → Get taskId
Poll Status → Wait for success
Download → Get video URL

That's it. Three calls, one video.

🚀 Quick Start

git clone https://github.com/fracabu/kie-ai-app
cd kie-ai-app
pip install -r requirements.txt

Add your API keys to .env:

KIE_API_KEY=your-key
OPENROUTER_API_KEY=your-key  # for Claude storyboards

Run it:

python video_generator.py

📦 What's in the Repo

Script	What it does
`app.py`	Streamlit web dashboard
`video_generator.py`	Full pipeline with Claude storyboards
`genera_da_storyboard.py`	Generate from existing storyboard JSON
`merge_videos.py`	FFmpeg merge utility

🎨 Available Models

Model	Use Case	Cost
`sora-2-text-to-video`	Text to video	$0.15
`sora-2-image-to-video`	Animate images	$0.15
`sora-2-pro-*`	Higher quality	$0.75+

📊 My Results

I've generated 50+ videos for various projects.

Metric	Value
Total spent	~$15
Videos generated	50+
Traditional cost	$2,500+
Savings	$2,485

💡 Pro Tips

Write Cinematic Prompts
- Include camera angles, lighting, movement
- Specify "4K", "cinematic", "slow motion"
Watch for Gotchas
- n_frames must be string: "15" not 15
- URLs expire in ~24h — download immediately
Use Claude for Storyboards
- AI generates consistent scene prompts
- Adds only ~$0.03 to total cost

🔗 Links

🐙 GitHub: fracabu/kie-ai-app
🎬 Kie.ai: kie.ai
🤖 OpenRouter: openrouter.ai

🎬 TL;DR

OpenAI Sora: $20/month
Kie.ai Sora 2: $0.15/video
Full 60s promo: ~$0.63

Stop overpaying. Check out the repo and start generating.

Have you tried AI video generation? What tools are you using?

Drop a comment — I'd love to hear your experience!

If you found this useful, consider giving the repo a ⭐ and following for more AI tool builds!

From Warehouse to Code: My Self-Taught Journey into AI & Multi-Agent Systems

fracabu — Tue, 16 Dec 2025 12:28:40 +0000

Hey everyone, I'm Francesco 👋

Today, I want to share a story that isn't about the latest framework or a complex architectural pattern, but about a personal journey – one that took me from the demanding environment of a warehouse floor to the fascinating world of AI and multi-agent systems.

If you're reading this, chances are you're a developer, or aspiring to be one. And maybe, just maybe, you've felt the pressure of not having a "traditional" background – no Computer Science degree, no expensive bootcamps. Well, you're not alone. My story is living proof that curiosity, persistence, and a passion for building can truly beat credentials.

📦 The Starting Line: Life in the Warehouse

For a significant part of my professional life, my days were defined by logistics, inventory, and the physical demands of a warehouse. It was an honest job, one that taught me discipline and the value of hard work. But deep down, there was a growing itch, a nagging curiosity about how the digital world worked, how software was built, and what truly made things "smart."

I watched, I wondered, and eventually, I decided to dive in. The idea of transitioning into tech felt daunting, almost impossible, especially without a formal background. But that curiosity was a powerful motivator.

🎓 My University: Documentation & AI as My Teacher

I didn't have the luxury of a university curriculum or a structured bootcamp. What I did have was access to the internet, an insatiable desire to learn, and two incredibly powerful allies: comprehensive documentation and AI assistants.

My learning approach was simple yet effective: learn by building.

Whenever I encountered a new concept, a new language like Node.js, TypeScript, or Python, I wouldn't just read about it. I'd immediately try to build something, even if it was a tiny, broken prototype.

🤖 How AI Became My Personal Tutor

AI assistants like Claude AI and Google Gemini weren't just chatbots for me; they were my personal tutors, available 24/7:

Use Case	How It Helped
Explaining Concepts	When documentation seemed opaque, AI explained it in simpler terms with examples
Debugging	Paste code + error → get pointed in the right direction
Code Generation	Quick boilerplate and prototypes to focus on the why
Best Practices	Instant, distilled knowledge on demand

This wasn't about letting the AI code for me, but about having an intelligent pair-programmer and a patient teacher always at my side.

🤖 Diving Deeper: AI & Multi-Agent Systems

As I gained confidence in fundamental programming, my fascination naturally gravitated towards Artificial Intelligence. Specifically, the idea of multi-agent systems captivated me — specialized AI agents collaborating to solve complex problems, mimicking a team of human experts.

Featured Projects

fracabu / Agent-PayMind

PayMind is an AI payment reminder system that autonomously analyzes invoices, generates personalized messages, and handles customer responses using Claude AI agents.

💰 PayMind

AI-Powered Payment Reminder System

Italiano • English

🇮🇹 Italiano

Panoramica

✨ Caratteristiche

Funzionalità	Descrizione
🤖 3 Agenti AI Specializzati	Team di agenti che collaborano per gestire l'intero workflow
📊 Analisi Intelligente	Identifica fatture scadute, calcola priorità e segmenta clienti
📧 Multi-Canale	Genera messaggi per Email, SMS e WhatsApp
💬 NLP Avanzato	Analizza risposte con intent recognition e sentiment analysis
🎨 Dashboard Moderna	Interfaccia Next.js per visualizzare workflow in tempo reale
🔌 Multi-Provider AI	Supporta Anthropic, OpenAI, OpenRouter e Google Gemini
🗄️ Database SQLite	Persistenza dati con Prisma ORM
🌙 Dark/Light Mode	Toggle tra tema scuro e chiaro
🌍 Bilingue	Supporto completo Italiano e Inglese
📊 Grafici Interattivi	Dashboard con

…

View on GitHub

fracabu / tube-forge-agents-team

Multi-agent AI system for YouTube content production. 6 specialized Claude Code agents automate the entire video pipeline from market research to publishing.

TubeForge Agents Team

AI-Powered YouTube Content Production Pipeline

Multi-agent system for automated video content creation

🇬🇧 English | 🇮🇹 Italiano

Agent Architecture

Tech Stack

🇬🇧 English

Overview

TubeForge is a multi-agent orchestration system built with Claude Code that automates the entire YouTube video production pipeline. It coordinates 6 specialized AI agents to produce high-quality content for Italian small business owners interested in AI automation.

Features

Sequential agent pipeline - Each agent reads prior outputs and produces standardized files
Structured asset generation - 10 numbered files per video (scripts, prompts, metadata, checklists)
Bilingual support - Italian primary content, English metadata
Market-validated niche - AI automation for small businesses
Complete workflow - From research to publishing

Target Audience

Italian small business owners (28-55 years) seeking practical AI automation solutions. Content is expert, ROI-focused, with 12-15 minute tutorial format.

🇮🇹 Italiano

Panoramica

TubeForge e un sistema di orchestrazione multi-agente costruito con Claude Code…

View on GitHub

fracabu / sellsnap

SellSnap is an AI-powered web application that provides instant, professional appraisals for second-hand items, including price, description, and pre-filled data for selling platforms.

SellSnap

AI-Powered Second-Hand Appraisal

Instant valuations for used items with Gemini AI

🇬🇧 English | 🇮🇹 Italiano

🇬🇧 English

Overview

A web application using AI to provide instant appraisals for second-hand items. Snap a photo, get a professional valuation with description, price, and ready-to-sell fields for Vinted, eBay, Subito.

Features

Multi-Category AI Analysis - Antiques, furniture, electronics, fashion, music, watches, jewelry
Price Estimation - Market research and comparisons
Authenticity Checks - Counterfeit and legal risk detection
Multi-Platform Fields - Pre-filled for Vinted, eBay, Subito
Interactive Chat - Follow-up questions on appraisals

Quick Start

git clone https://github.com/fracabu/sellsnap.git
cd sellsnap
npm install

# Create .env.local
GEMINI_API_KEY=your_key

npm run dev

Access at http://localhost:5173

🇮🇹 Italiano

Panoramica

Un'applicazione web che usa l'AI per fornire perizie istantanee di oggetti usati. Scatta una foto, ottieni valutazione professionale con descrizione, prezzo e campi pronti…

View on GitHub

These projects aren't just code; they're my proving ground, my way of understanding and pushing the boundaries of what AI can do when orchestrated effectively.

🚀 Real-World Apps: From Idea to Production

Beyond npm packages and AI agents, I've shipped 35+ web applications to production. Here are some highlights:

AI-Powered Tools

App	What it does	Try it
Remember Me	Voice recordings → structured reminders (Gemini AI)	Live Demo
SEO Analyzer	On-page SEO analysis powered by Gemini	Live Demo
Promptify	Explore 28 Prompt Engineering frameworks	Live Demo
SellSnap	AI instant appraisals for second-hand items	Live Demo
GitHub Repo Describer	Auto-generate repo descriptions	Live Demo
Prompt Craft	Generate structured C.R.A.F.T. prompts	Live Demo

🏠 Hospitality Tech (Solving Real Problems in Rome)

Living in Rome, I noticed gaps in the vacation rental industry. So I built tools to solve them:

Tool	Problem Solved
Overbooking Shield	Prevents double bookings via iCal analysis
Host Reservation Panel	AI assistant for Airbnb/Booking hosts
Tassa Soggiorno	Auto-calculates Rome's tourist tax

Each started as a problem I wanted to solve, built with the "learn by building" approach.

📦 Open Source: Fastify Plugins

Another crucial part of my journey has been building in public. My npm packages for the Fastify ecosystem:

View all 7 npm packages

Package	Description	Link
@fracabu/fastify-api-key	API key authentication with scopes	npm
fastify-rfc9457	RFC 9457 compliant error responses	npm
fastify-cache-control	Declarative HTTP Cache-Control headers	npm
fastify-webhook-verify	Webhook signature verification (Stripe, GitHub)	npm
fastify-i18n-plugin	Internationalization for Fastify	npm
@fracabu/env-checkup	Environment variable validation CLI	npm
node-test-expect	Jest-like assertions for Node.js native test runner	npm

Creating and maintaining these packages taught me the full lifecycle of production-ready tools. It's rewarding to know other developers are using what I've built.

💡 Advice for Fellow Self-Taught Developers

If my story resonates with you, here are a few takeaways:

Embrace Curiosity — It's your most powerful asset
Build, Build, Build — Practical application beats theory
Leverage AI Smartly — Use it as a mentor, not a replacement
Dive into Documentation — Your most accurate textbook
Share Your Work — Building in public leads to opportunities
Solve Real Problems — The best projects scratch your own itch

🔮 What's Next?

My journey is far from over. Currently diving deeper into:

🔄 Multi-agent orchestration patterns
🐍 Machine Learning with Python
🌍 Contributing to larger open-source projects

🤝 Let's Connect!

🐙 GitHub: @fracabu
📦 npm: fracabu packages
💼 LinkedIn: Francesco Capurso

What's your unique journey in tech? Whether you're self-taught, career-switching, or just getting started — drop a comment. I'd love to connect!

If you found this inspiring, consider giving it a ❤️ and following for more content on AI agents, open-source development, and the self-taught developer life.