Mission & Roadmap
Frontier AI, held to a standard, in the service of learning
My Weird Prompts explores how frontier AI models can deliver high-quality, informative educational content. It does this through a rigorous implementation of agentic AI: research grounding, a version-controlled pipeline and documented decisions. The result is up-to-date, evidence-based content that helps listeners across many areas of life, from technology to parenting, and that deepens our understanding of the world and our part in it as citizens.
Why it matters
Most people have more questions than time to research them. A well-built AI pipeline can take a single question and return a grounded, half-hour exploration of it, at a cost of cents. Published openly, every one of those explorations becomes a public resource rather than a private chat that disappears.
The hosts are a sloth and a donkey, and the show does not take itself too seriously. The engineering behind it does. Bootstrapped and built by one person, the project has already published 5,932 episodes and 2,664 hours of audio, and it sets out to show what this approach can achieve when it is done carefully.
How we hold it to a standard
Grounded
Every script is written from research gathered for that episode, not from the model's memory alone.
Version-controlled
The pipeline, its prompts and every design decision live in version control, and model changes are logged episode by episode.
Correctable
Errors are fixed in public under a stated corrections policy, and anyone can report one.
Open
The process is documented and the archive is published as open datasets, with a public MCP server for programmatic access.
Honest about AI
The show never hides that it is AI-generated. Listeners always know which parts are human and which are synthesised.
Efficient
A full episode costs well under a dollar to produce, which is what makes a catalogue of this size sustainable.
What has been built so far
The project started as a proof of concept. It has gone well beyond that.
- An end-to-end agentic pipeline that takes a spoken question through research, scriptwriting, review, voice synthesis and publication. Technical white paper
- Open datasets of the full episode archive and transcripts on Hugging Face. Research uses
- A public MCP server that lets AI agents search and cite the catalogue. MCP server
- A mirror network with checksummed archive volumes, so the catalogue does not depend on any single platform. Mirrors
- Custom AI models built to solve specific production problems, starting with Hebrew pronunciation.
Where it is going
Until now nearly all the effort has gone into the pipeline itself. Analytics were added only recently, and there has been almost no time for curation or outreach. The roadmap is about bringing the same rigour to organising the content, reaching new audiences and moving past the limits of the podcast format.
Organise and curate the catalogue
Graph-based knowledge exploration
With 5,932 episodes, the catalogue is large enough to map. The aim is to use graph databases to build immersive environments where self-learners move from one topic to the adjacent topics they did not know to look for.
Age-appropriateness review
A retrospective pass over the whole back catalogue to identify and certify the episodes that are suitable for younger listeners, opening the archive to validated educational use.
Curation and editorial structure
Organising the archive into curated collections, learning paths and series, so that its depth is visible rather than buried in volume.
Reach more people
Independent mirrors
The full catalogue is distributed through checksummed, independently run mirrors, so that no single platform decides whether it survives.
Access under internet censorship
Reaching listeners in regions with restricted or censored internet, including distribution through censorship-circumvention networks such as CENO.
Offline listening appliance
A self-contained device that carries the catalogue and plays it without a connection.
Multilingual episodes
Every episode starts life as a script, so it can be produced in other languages. Work on mixed-language speech has already begun with Hebrew.
Move beyond the podcast format
Choose your own voices
Each episode is a transcript before it is audio. Listeners will be able to pick from preset voices and hear the same episode their way, without the constraints of the podcast spec.
Audio to video
As the cost of video and animation falls, the existing bank of topics becomes source material for video and interactive content, starting with stock footage layered over the audio.
Raise quality with our own models
Purpose-built small models
Where general-purpose models fall short, the project trains its own. The first is a token classifier that detects Hebrew words written in Latin letters so they can be pronounced correctly.
Models trained on the catalogue
Small classifiers and a text normaliser for speech, trained on the episode archive, to replace general-purpose API calls with predictable, near-free models inside the pipeline.
Fine-tuning on the transcript dataset
Using the open transcript dataset to fine-tune a script-writing model for the show.
Work with us
We are looking for organisations that share these aims, in education, open knowledge, digital access and AI, to help with compute, credits, expertise or sponsorship.