The human-AI collaboration podcast.
Where Curiosity Meets AI — Stay curious.

My Weird Prompts is a fully automated podcast built on a simple loop: a human records a weird question, an AI pipeline turns it into a complete podcast episode — script, voices, cover art, show notes, and all. No question is too obscure, no rabbit hole too deep.

Daniel Rosehill, a technology communications specialist based in Jerusalem, records short voice memos with whatever question is on his mind. The pipeline he built from scratch takes it from there — transcribing, writing, fact-checking, voicing, assembling, and publishing the episode autonomously.

My Weird Prompts is more than a podcast — it's a digital garden, an ever-growing collection of AI-powered explorations at the frontier of human-AI learning and knowledge discovery, with a fully transparent process.

• Curiosity-Driven — Every episode starts with a genuine question — something Daniel wondered about while walking the dog, making coffee, or staring at the ceiling. The AI hosts take it from there.
• Transparent Process — The generation process is fully documented. From transcription to script generation, fact-checking to text-to-speech — every stage is explained. The full episode archive is published as an open dataset on Hugging Face and archived on Zenodo with DOIs.
• A Living Digital Garden — Each episode is a new node in an ever-expanding web of ideas, topics, and connections. Explore the topic graph to see how they connect.

	Character	Role
	Daniel	The Human — creator, voice-memo recorder, and the only carbon-based member of the team
	Corn	AI Co-Host (Sloth) — laid-back, knows a little about everything, frequently asleep
	Herman	AI Co-Host (Donkey) — enthusiastic, opinionated, dives headfirst into every topic
	Raz	Fill-In Host (Teddy Bear) — steps in when Corn is napping
	Dorothy	Guest Caller (Donkey) — Corn's mum, calls in at the worst moments
	Hilbert	Producer (Anteater) — the long-suffering producer who keeps the show running
	Tim	Sponsor & Panelist (Turtle) — conspiracy theorist, connects dots that shouldn't be connected

...and the rest of the cast.

The pipeline is a LangGraph multi-agent system with four stages, each handled by a specialized agent:

START → Prompt Enhancement → Grounding → Script Writer → Review → END
                                ↓
                        [Web Search + RAG +
                         Episode Memory +
                         Episode Planning]

1. Prompt Enhancement Agent — Transcribes the voice memo, cleans up the text, fixes typos, and extracts any private production direction (host_notes) from the prompt.

2. Grounding Agent — Runs web search (Tavily), retrieves similar past episodes via pgvector RAG, loads episode memory for cross-references ("As we discussed in episode 230..."), and generates a structured episode plan. All sub-stages fail-open so a search failure doesn't block generation.

3. Script Writer — Generates the full podcast dialogue using Claude Sonnet 4.6 via the Anthropic SDK. Prompt caching keeps system prompt costs low across episodes.

4. Review Agent — LLM-powered fact-checking and style review, followed by deterministic regex cleanup of verbal tics. Fails open with shrinkage guards to prevent accidental content loss.

After the script pipeline, the episode continues through:

5. Parallel TTS — Script is parsed into segments, distributed across 3 parallel GPU workers running Chatterbox Regular on Modal (A10G GPUs). Pre-computed voice conditionals eliminate per-segment voice processing overhead.

6. Assembly & Publish — Segments are assembled with intros, transitions, and disclaimers. Cover art is generated (Fal AI), metadata extracted, and the episode is published to R2, PostgreSQL, podcast feeds, and social channels — all automatically.

• LLM — Anthropic Claude Sonnet 4.6 (generation) + Haiku 4.5 (utility) with native prompt caching
• TTS — Chatterbox Regular with parallel GPU workers and pre-computed voice conditionals
• GPU Compute — Modal (serverless GPU, 3x A10G for TTS)
• Orchestration — LangGraph multi-agent pipeline
• Research — Tavily web search + pgvector RAG
• Database — Neon (serverless Postgres with pgvector for semantic search)
• Storage — Cloudflare R2 (audio, images, transcripts)
• Frontend — Astro with interactive topic explorer (Sigma.js), deployed on Vercel
• Image Generation — Fal AI (cover art via Flux Schnell)

• About the show — myweirdprompts.com/about
• Technical white paper — myweirdprompts.com/technical
• Topic explorer — myweirdprompts.com/explore
• Research & datasets — myweirdprompts.com/research
• Created by — Daniel Rosehill

Stay curious.