How It Works

From voice memo to finished episode in about a quarter of an hour, live on the site the moment it publishes — no editing, no mixing, no manual steps.

The Pipeline

Every episode of My Weird Prompts follows the same automated journey. I submit a prompt — by voice or text via the send app, or straight from an AI agent over MCP — and the pipeline handles everything else.

01

Submit a Prompt

I send a topic to the pipeline — via the send-app at send.myweirdprompts.com 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, news briefings, interviews, and more — each with its own format and cast. Each job waits 60 seconds before its first paid call, so a mistaken submission can be cancelled for free. A prompt that repeats one already in the queue, even reworded, is refused.

02

Transcribe & Research

Voice recordings are transcribed in one pass by Gemini Flash Lite (with Whisper as the fallback), then a research agent investigates the topic, checks past episode coverage with semantic search, and an episode plan is built with key facts, context, and structure.

Research combines Exa web search (hour-level freshness) with primary-source tools loaded by topic (Hugging Face, GitHub, arXiv and full papers, Wikipedia, PubMed, Sefaria), plus a vector search over past episodes and the hosts' memory for cross-references.

03

Generate the Script

A large language model writes a full dialogue script featuring Corn and Herman, with an on-air turn from Hilbert, the show's producer, in every episode. Every LLM stage, including the script, runs on DeepSeek V4.1 Flash via the DeepSeek API.

A standard script runs 3,700–4,300 words, producing an episode of about 25 minutes; short (15 minutes) and extended (30 minutes) tiers sit either side. Left to itself, the planner picks a length between 20 and 30 minutes, with deeper topics getting the longer slot.

04

Polish

A deterministic polish pass cleans the script — no second LLM is involved. It strips verbal tics, rotates overused phrases to synonyms, and names related past episodes in the sign-off, then optimizes the text for speech.

Because the polish is code rather than a model, it is repeatable: the same script always gets the same cleanup, and it never rewrites or truncates the hosts' argument.

05

Generate Audio

Each line of dialogue is sent to Chatterbox TTS running on three parallel L4 GPU workers via Modal. Each character has a unique voice clone — Corn's lazy drawl, Herman's eager bray, Hilbert's deadpan, and a clone of Daniel's own voice for the AI disclaimer.

Numbers and symbols are converted to spoken words first. TTS runs across 3 parallel L4 workers with pre-cached voice embeddings and a CUDA-graph fp16 decode — about 18 cents of GPU time per episode (median, mid-October 2026). Safety checks abort if more than 20% of segments fail.

06

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 Telegram — all automatically, with no site build in the way.

A script below the length floor for its planned duration, or with fewer than 10 lines of dialogue, is rejected before any GPU time is spent, so an incomplete episode is never published.

By the Numbers

~13 min Median pipeline time
~$0.26 Median cost per episode
6 Pipeline stages
0 Manual editing steps

Want the Full Technical Details?

The complete pipeline architecture, cost analysis, and lessons learned are documented in the technical white paper.