I Know Kung Fu

vibecoding claude transcriptions podcasts llm

My history with computers goes back over 40 years, and there have been four defining moments in my relationship with this technology: my first computer (an Atari ST in 1986), my first time on the open internet (1994), my first conversation with ChatGPT (2023) – and one that just happened.

Finding Claude

After ChatGPT launched, I explored various AI assistants with growing curiosity. When I discovered Claude, something felt different. While competitors chased flashy features, Anthropic acted like the grown-up in the room — thoughtful, deliberate, and focused on building AI that’s genuinely beneficial for the user.

Claude quickly became more than a chatbot to me. It helped set up my Proxmox cluster, debug network issues, and write scripts. I soon realized: Claude doesn’t just explain how things work – it actually collaborates.

So I subscribed. Then I renewed for a year. I chatted, I copied, and I pasted. And right before New Year’s, I installed Claude Code.

And that changed everything.

cast2md: From Idea to Production in Two Weeks

In January 2026, I built cast2md – a full-stack podcast transcription service. Two weeks, 431 commits. Running in production on my own infrastructure.

I’m by no means a professional developer – I’m a 52-year-old IT guy who spent years in enterprise consulting and management and wrote my last serious piece of code over 20 years ago. But with Claude Code, I shipped a real product. Concepts I only knew from theory suddenly made sense through building.

The idea was simple: I wanted to turn podcast RSS feeds into a searchable, LLM-ready transcript library. What came out of it was anything but simple:

  • Transcript-first workflow — cast2md checks for publisher-provided transcripts (Podcasting 2.0) and Pocket Casts before falling back to local Whisper transcription. No wasted compute.
  • Hybrid search — full-text search via PostgreSQL’s tsvector combined with semantic search via pgvector embeddings, fused through Reciprocal Rank Fusion. You can search by keyword, by meaning, or both.
  • MCP server for Claude — a full Model Context Protocol integration that lets Claude search transcripts, summarize episodes, and explore topics across an entire podcast library.
  • Distributed transcription — remote worker nodes on M4 Macs, GPU PCs, or RunPod pods that poll for work and transcribe in parallel. Scales from a single laptop to a fleet of GPUs.
  • Web UI, REST API, and CLI — a dashboard for managing feeds and episodes, a complete API for automation, and a CLI for everything else.

The stack: FastAPI, PostgreSQL with pgvector, faster-whisper and mlx-whisper for transcription, sentence-transformers for multilingual embeddings, Docker for deployment. All self-hosted, all open-source.

The source code is on GitHub, and the project page has the full documentation.

The “I Know Kung Fu” Moment

It was suddenly like knowing a new language — a direct connection between intent and implementation. My personal “I know kung fu” moment. I’d describe what I wanted, Claude would write the code, I’d test and adjust, and the loop just kept going. Not copying snippets from Stack Overflow. Not fighting with documentation. Actually building.