PodcastAI: Topic to Published Podcast — Fully Automated Pipeline
Joey Musselman · February 15, 2026
Producing a podcast episode typically costs $500-2,000 and takes hours of recording, editing, and post-production. I built a system that generates a complete episode — research, script, audio, transcription, and publishing — from a single Telegram command. Episodes are live on Spotify and Apple Podcasts.
Challenge
The pipeline needed to chain 9 distinct steps reliably: topic research with Claude API, two-host script generation with multi-step prompt chains, high-quality voice synthesis for both hosts, audio stitching and transcoding, transcription with word-level alignment, metadata generation, cloud upload, RSS publishing, and episode archiving. Each step could fail independently, and the system needed to report progress, support cancellation mid-pipeline, and optimize token usage across multiple Claude API calls.
My Role & Contribution
Sole Developer — AI Pipeline Architecture, Prompt Engineering & Deployment
- Designed a 9-step pipeline: research → script → TTS → stitch → transcode → transcribe → metadata → publish → archive
- Engineered multi-step Claude API prompt chains for research, script generation, and metadata with token optimization across stages
- Implemented high-quality dual-voice synthesis using Kokoro TTS (ONNX model)
- Added WhisperX transcription with word-level alignment and speaker diarization
- Built job queue with Telegram progress updates, cancellation support, and episode numbering persistence
- Published to Spotify and Apple Podcasts via RSS distribution through Cloudflare R2
- Deployed as systemd service on Linux with SCP deployment pipeline
Key Technologies
Impact
Generates complete podcast episodes end-to-end from a single Telegram command. Episodes include two distinct AI hosts, professional-quality audio, full transcriptions with timestamps, and automatic publishing to Spotify and Apple Podcasts via RSS. Multi-step prompt chain demonstrates production Claude API integration with token optimization across research, script, and metadata stages.
View live →