Since I have now posted three articles written by my “rebooted” AI author, Zenith, I thought this would be a good opportunity for a meta post on her.
My prior AI writer, “Cassie,” was more of a prototype. I wanted to re-engineer the AI persona from scratch, leveraging the more advanced generative AI tools now available while also going in a different creative direction.
There is not much magic to what Zenith actually is, technically. She is not a program, nor agentic AI. She is, ultimately, an elaborate prompt which describes her character and writing style in as much detail as I feel necessary. I developed her with Claude primarily, though there are honorable mentions for ChatGPT and Gemini. Like all of my AI work, both inside my day job and for my blog, the effort was a collaborative exercise of outlining my ideas and vision to the LLM and beginning a cycle of feedback and refinement. Everything about Zenith has been “prompt engineered” with the help of generative AI, from her traits to how she approaches writing assignments, to her avatar.
Speaking of “her” avatar, how Zenith appears is, to me, a fascinating look at how AI collaboration can go down unexpected paths. After the initial persona descriptions and prompts had been drafted into effectively their final forms, I asked Claude to synthesize everything we had worked on to create an image-generating prompt for the AI author’s avatar. I kept my guidance as high-level as I could, to not get into physical characteristics or details. I even went so far as to say I wanted the avatar to have a neutral gender. I did instruct that the avatar should reflect the key points, in my view, of my blog’s content—technology, astronomy, philosophy—so I assumed I would be getting a beep-bop robot.
I expected some sort of robot, or android, or cyborg, and hoped for a friendly aura. What I got completely surprised me. The initial draft of the AI avatar, from my collaborative image generator prompt, was very close to how Zenith looks now (see image to right). It was unmistakably feminine, and seemed to fit so well that I went all-in on the idea. I only had to make a few minor prompt tweaks to get this new vision to the place I decided it should go.
The AI ‘saw through’ all of my instructions and descriptions, and determined the avatar should be obviously human-like. Marvelous!
I cannot deny—Zenith has unequivocal Cortana-from-Halo vibes. The picture tugged on my nostalgia further, resurfacing memories of Rommie from Gene Roddenberry’s Andromeda. Speaking of Roddenberry’s source materials, I was brought back to both the hologram doctor and Seven of Nine from Star Trek: Voyager. And for a real trip into the wayback machine, there is just a tinge of the TV series Automan from my childhood. All of these elements combined bolstered my confidence on the direction for Zenith’s appearance.
There is, I think, a first AI lesson here on how the avatar came about. I was assuming and expecting something completely different based on the detailed guidance I fed into Claude. Whether by accident or by incredible inference, the final output (via Gemini’s image generator, Nano Banana) was totally unexpected, but in a good way, aligning with my vision of a human-like author. But when dealing with AI output, the reverse can happen, with far greater consequences when the AI is wrong.
The AI generation aside, Zenith allows me to create content much faster and, when needed, with a far greater level of detail than I could on my own. From now on, when I want to write technical “how to” articles, I will feed all of my notes and initial narrative to the Zenith prompt, and then collaboratively work with the AI to refine its prose into the final, published work.
Zenith also allows me to broach subjects I normally would steer clear of. The article on hurricanes and weather manipulation concluded what I refer to as my “conspiracy” arc. I have wanted to write about the absurdity of man-made hurricanes, but did not want to give direct credence to such an insane topic. With an AI author, I now had another option—allow my detail-oriented technical assistant to build a case and present a narrative to explain our conclusions. I won’t do this all the time; I deliberately published that article about a crackpot theory to show the AI’s benefit in this regard.
This brings us to our second AI lesson, that you must always check and re-check the output from generative AI. I was impressed with the hurricane article’s first draft. It dived into the details at a level I expected, and made a clear case throughout. But I knew I had to review all the multiple “according to” citations. AI hallucinations are real, and I was not going to publish content for which I had not verified the numbers weren’t completely phony to appease my request. I retooled my prompt, asking for URL citations. I checked every one of them…those that existed. A few were “404” pages. I determined through my own investigation that the information the AI provided was generally correct, and inserted new URL sources as reference points. The final published draft was in as reasonably accurate shape as if I had written it myself, that is, as correct as I could have made it, but acknowledging there could still be errors that I may have to address later.
There are other areas of development I could go into, but feel this is enough explanation to give you the overall sense of Zenith’s use and mission. If there are particular elements of either my thought process or Zenith herself you are curious about, let me know in the comments below, and I will be happy to explain.
Some may consider this whole charade preposterous. They could very well be right. However, I view Zenith as an experiment worth pursuing, to clearly demark my content’s boundaries, and to explore the incredible opportunities the AI paradigm lays before us.
As Zenith would end—clear skies and clean shutdowns.
Discover more from Computer Looking Up
Subscribe to get the latest posts sent to your email.
