Making Tome of Stars:
Process and Meaning

In constructing the sonnets in Tome of Stars, I combined my previous approaches for writing structured poetry with several modern tools, including online thesauruses, online rhyme dictionary/generators, and generative AI. In the past, all the way back to spiral notebooks, I would write out in prose what I wanted to say when producing a structured poem (e.g., a sonnet). First loosely, in a paragraph or two. Then I would cull that text down to the essence of what I felt was important, beginning a process of substituting imagery, symbolism, and metaphor in place of the more “prosaic” expressions. At this stage, I would start producing lines, moving toward meter and rhyme. I would then go through iterative steps of refining the lines to fit the constraints and rules of the poetry style I was writing in (obviously this is for highly structured poetry, as free verse did not require the same type of effort). I would often write multiple versions of a line, sometimes slight variations, sometimes five or six very different takes on what I was trying to say, with different vocabulary, imagery, etc. Part of the process was just generating options for expression and integration into the formula of the structured verse.

With generative AI, I was able to augment this process and create many more options at the expressive level. For each poem in the collection, I would begin with one of the following (also on a spectrum): (1) a sonnet draft at various stages of refinement, (2) prose lines organized in stanzas to guide the AI, or (3) sometimes just the prose from my journals, letters, and occasionally book excerpts for the later “companion” volumes to Tome of Stars. I would then have LLMs create highly defined, structured iambic pentameter verse from these provided writings (as best they could), instructing it to stay close to the meaning and organization I had (in the case of more developed writing) and to produce different sonnets varying in perspective, vocabulary, imagery, etc. This is similar to my own “by-hand approach,” but automated. For each poem, I might use 2–3 different generative AIs and ask them to output 5–10 sonnets from my initial writings. I would select from those.

I was using ChatGPT, Claude.ai, Grok, and Gemini as they existed in the 2023-2024 (e.g., GPT-4o, o1-preview, Gemini 1.5, Claude 3.5 Sonnet, Grok-2). These would return verse at various levels of competency, not infrequently failing to fit the meter and rhyme, and commonly producing (to my ear) painful verse, and certainly a lot of verse that did not fit with the intent and style I desired. Yet there were often elements I liked and found useful, especially when I produced a large sample size.

Next, I would have my original writings at the top of a text page and paste in the generative sonnets beneath. Line by line, using my original writings, I would hunt through the generated material for expressions, words, phrases, sometimes entire lines, that transmuted what I said into something that captured my intent and vision for the verse. This might mean I would chop and stitch together pieces of quite a large number of different synthetic poems whose verse otherwise was not used. When I felt I had something that was working to capture and express my source writings, I would then move to a completely manual portion of the process.

Here I would go word by word, line by line, and rework, substitute, completely rewrite, and in some cases leave just as it was, the interwoven building blocks from the synthetic output. Parts of this I might feed back into LLMs for refinement or generating many variants, which then would enter (or not) into the manual rewrite stage again. This would be a cycle of editing, regeneration, and rewriting until I had taken the text as far as I could go toward my vision.

Next, I would drop the finalized sonnet back into generative AI to (1) analyze it from both a literary and grammatical/syntactical viewpoint in a kind of a virtual poetry critique, leading to more rewriting, and (2) undertake a final polish to punctuate the verse according to English poetry norms (because I have never been able to truly grasp what seem to me the often quite arbitrary and inconsistent rules for punctuation, which other languages often disagree about quite strongly). I would also read the poems out loud and sometimes catch issues (to my ear).

I then subjected my entire collection to statistical analyses of word/phrase usage, frequency, clichés, etc., trying to find those vices in which my mind would subconsciously indulge (I used Autocrit, a writing-analysis service heavily employed in analyzing my novels pre-AI). Another round of edits would seek to purge some of these problems (and also inspired the creation of two new poems based on word frequency alone: Dance, Stars! and Frequency Modulated).

Finally, I hired freelance editors (e.g., Reedsy) who brought a professional eye to the work. Their comments required a final round of manual edits. Of course, writing never ends, it’s just abandoned at some point. I still find errors, both basic ones as well as “errors of art”. A good example of the latter is in the poem , where I realized I should have chosen a single word differently (“sign” instead of “sight”) for several impactful artistic reasons. But, I have ended the project, so it is what it is, imperfections and all.

As brief and informal examples, I’ve linked below two poems that illustrate a little of the process with two starting points of my writing that lie on different ends of the spectrum. Last Letter (a rather late sonnet in Item 5: Quixotosis) begins from a draft sonnet with organized quatrains and a couplet, already mostly in iambic pentameter, but lacking the rhyming and some internal text marked by placeholders. In contrast, Princeton (from Item 3: Sporogenesis) began with an excerpt from an email I sent the Stargirl in 2005.

In all, the time spent to create a book of 100 sonnets (about 10,000 words) was equivalent to the time I took to write my novels (80–100K-word books). Because I have a day job that keeps me housed and fed, as with my novels, I found this time in odd hours of the day (usually at night). I would also note that the use of generative AI didn’t speed up the process. In the end it was basically similar to the time it took me to write sonnets before these tools became available. However, like with other aids (thesaurus, rhyming dictionaries, etc.), at each stage I had many more choices for ideas and language than my limited mind could produce. This was a strong motivation to use these tools, as my goal was to try to build the best verse I could in this long and absurd love letter about and to the Stargirl. Choosing from among the generated options from my initial writings did not take much longer than how I would have done it generating my own (less numerous) variants, but there were more options, and the time gained and lost basically balanced out. 

Always guiding my efforts were the very real events and emotions underlying this story of transformative love, loss, and grief. Everything I did was to make the verse convey, as closely as possible to my vision and taste, the churning universe of feeling within me about my beloved. From journal entries to rhyme generators to ChatGPT/Grok/Claude.ai to statistical analyses, each of these elements served as the marble, chisels, rasps, files, polishing stones and paper that I used to produce earnest, honest creation.

Of course, I failed. We all do in trying to capture the uncageable mystery of feeling within us. But this was my writing process.