Clown Cast
A podcast I built for myself. AI researches and writes the episodes. Most of the work is in checking what it says.
Daily episodes cover whatever I'm curious about. Study series, sourced by hand, cover what I'm learning or building, and are written only after the work happened. Listen on Spotify.
The checks
- Every claim is tagged in the hand-sourced notes: verified, unverified (the hosts must say so) or unknown (it stays open on air).
- Every script gets a number check. It flags spoken percentages that aren't in the research, and pairs like 2% and 98% that might be one number turned inside out.
- Corrections go on air, in a later episode, with the better version.
What the checks found
An episode hedged one shaky statistic and then quoted a scarier one, a 98% failure rate that was really 2% of plants saying they had fully met their goals. Mark Graban had traced it on his blog. An audit of 106 episodes then found 22 of 245 spoken percentages missing from the research. The real problem was upstream: the prompt asked for “surprising statistics”. That line is gone, and every number now has to say who measured it.
Auditing the prompt itself found that its examples of good metaphors had become defaults, and that the classifier counting the show's themes read only the last 20 lines of each script. Both were fixed on air.