How It Works
From voice memo to published podcast episode in under 30 minutes — no editing, no mixing, no manual steps.
The Pipeline
Every episode of My Weird Prompts follows the same automated journey. Daniel submits a prompt — by voice, text, or Telegram — and the pipeline handles everything else.
Submit a Prompt
Daniel sends a topic to the pipeline — via the MWP Recorder app on his phone, the Telegram bot, or the MCP admin server in Claude Code. Prompts can be voice memos, text topics, or even pre-written scripts.
The pipeline supports over a dozen episode types — standard discussions, panels, roundtables, debates, news briefings, interviews, and more — each with its own format and cast.
Transcribe & Research
Voice recordings are transcribed using AI, then a grounding agent researches the topic via web search, checks past episode coverage with semantic search, and builds an episode plan with key facts, context, and structure.
Research includes Tavily web search, pgvector RAG against past episodes, and episode memory for cross-references.
Generate the Script
A large language model writes a full dialogue script featuring Corn, Herman, and any other characters relevant to the episode. The script model is randomly selected from a pool of frontier models via OpenRouter for A/B testing.
Scripts typically run 3,000-6,000 words, producing episodes of 15-40 minutes.
Review & Polish
A cross-family LLM review pass refines the script — a model from a different family than the writer checks for accuracy, depth, and plan adherence. A deterministic cleanup pass then removes verbal tics and optimizes for text-to-speech.
Shrinkage guards reject edits that remove too much content. The review is additive-only — it never truncates.
Generate Audio
Each line of dialogue is sent to Chatterbox TTS running on parallel A10G GPU workers via Modal. Each character has a unique voice clone — Corn's lazy drawl, Herman's eager bray, Daniel's narration, and more.
TTS runs across 3 parallel GPU workers with pre-cached voice embeddings. Safety checks abort if more than 20% of segments fail.
Assemble & Publish
Audio segments are stitched together, silence is trimmed, and the final episode is assembled. Cover art is generated, metadata is extracted, and the episode is published to the website, podcast platforms, and social media — all automatically.
A duration gate rejects episodes under 10 minutes to prevent publishing incomplete content.
By the Numbers
Want the Full Technical Details?
The complete pipeline architecture, cost analysis, and lessons learned are documented in the technical white paper.