How I Actually Use AI
I discovered that Dear Future Overlords had two front doors—and they led to different shows
Dear Future Overlords is difficult enough to describe when everything is working.
It is a listen-first audio cartoon conversational between a human and an AI character. It moves through series, episodes, and parts. It is interested in agency, technology, creativity, and the deeply strange business of being human. It is not quite a podcast, not quite an audio drama, and definitely not a blog that occasionally remembers to contain sound.
The problem was that the two places a new person was most likely to arrive contradicted each other.
The website described the current DFO. Substack was still introducing an earlier version: “my favorite AI” taking notes for the robot uprising, a retired way of explaining the format, and subscription language that implied paid members received access to an archive the website said was not paywalled.
This was not one stale sentence. That sentence had roots.
Substack reused pieces of the old identity across the homepage, podcast metadata, subscribe page, automated emails, headers, footers, benefits, navigation, and section descriptions. The About page explained a version of the project that no longer existed. The actual About page and archive were hidden from navigation. The listen-first show did not show its podcast section on the homepage. Some episode numbers were duplicated or shifted, which meant episodes could appear in the wrong order in listening apps.
Meanwhile, the website had its own problems. The paths into the work were not as clear as the work itself. The homepage needed to center discovery around the series. The archive needed to reflect series, episode, and part instead of presenting one long undifferentiated wall. A listen-first project needed an actual Listen page. Contrast, keyboard access, old pages, broken annual-plan links, and stale content all needed attention.
This was less a redesign than an archaeological dig with deployment privileges.
I gave Claude the website and Substack as two connected jobs. On the website, it rebuilt the discovery structure, added the Listen page, reorganized the archive, created a shared content layer, improved contrast and keyboard access, repaired links, and removed retired pages along with 379 unused CSS rules. That version was built and deployed to production.
On Substack, Claude followed the old positioning everywhere it had propagated and replaced it with one consistent description of what the show is now. It rewrote the About page and five automated emails. It fixed navigation, subscription promises, podcast metadata, pricing inconsistencies, and episode numbering. It helped produce a podcast cover that could still be read at the 80-pixel size where podcast covers go to die.
The final audit recorded 35 changes, all checked against the public result rather than the comforting little message in an admin interface claiming something had saved.
We needed that verification rule because Claude also got things wrong.
It initially turned a browser-control failure into a confident explanation it could not actually support. An early audit miscounted how many publications recommended DFO. It missed one of two numbering gaps. It declared that episode numbering could not be fixed through the interface before discovering that the metadata panel had its own Save button. Substack's larger Done button could close the window while quietly discarding the change—a tiny user-interface decision engineered to create folklore about impossible problems.
We corrected the mistakes and kept them in the report. A clean history would have made AI look as though it understood the whole system on the first pass. It did not. The record now shows the current state and the places we got fooled, which should make the next audit better.
A visitor should never notice most of this work. Someone can arrive through the website, Substack, a forwarded email, a shared episode, or a podcast app and meet the same show. They can see how to listen, where to begin, what subscribing does, and what it does not do. The language no longer promises locked content in one place and open access in another.
Two front doors now lead into the same strange little universe.