I recently read a book called What We Can Know by Ian McEwan. McEwan calls it "science fiction without the science" — it opens from the perspective of a historian living in the future, studying a poem lost to time. The vanished work, A Corona for Vivien, is a love poem for the poet's wife (Vivien).
The book is not really about AI, but it made me ask: Just how good is Claude at writing poetry? Can AI write a decent sonnet? Where would it succeed, and where would it fall short? The result is an experiment and an analysis. The resulting poem is here. The note you're reading now explains how it came to be, and what I think of it. Claude has also written up its own thoughts; those are here.
The Form
A corona, or crown of sonnets, refers to a poetic structure built around a sequence of 14-line poems — the individual sonnets. There are a few accepted forms. Most forms link the sonnets by re-using the final line of the preceding sonnet as the first line of the next. A famous example is the 7-sonnet sequence La Corona by the English poet John Donne, which begins and ends with the line "Deign at my hands this crown of prayer and praise."
I decided to see if Claude could capture a form closer to the more ambitious structure described in What We Can Know. The following are the constraints of a heroic crown or sonnet redoublé:
- The first line of each sonnet (after the first) must match the last line of the preceding one. This is the standard rule for crowns.
- There must be 15 sonnets in a heroic crown.
- The final sonnet, number 15, is called the master sonnet, and has a special form. The first line of this sonnet is the opening line from the first sonnet; the second line is the line that opens sonnet 2, and so on. Thus, every line of sonnet 15 is the opening line from the corresponding earlier sonnet, for sonnets 1–14.
I decided to also constrain the last line of sonnet 14 to be the first line of sonnet 1, so that sonnets 1–14 alone were already a (non-heroic) crown. Why not?
I asked Claude to write a corona in which each sonnet was 14 lines, each line having 10 syllables in iambic pentameter (meaning that every other syllable is stressed, beginning with the second syllable of each line), with the rhyming structure ABBAABBA CDECDE. This is known as the Petrarchan form of a sonnet.
The Process
I used Claude Code, with the Opus 4.5 model, to create the corona. At a high level, I would ask for some writing or re-writing to be done, occasionally asking for notes or a snapshot to be saved to local files, and iterated through the terminal interface. I preferred this over the web interface because I have more direct visibility into sessions and control over the files.
I suspect it would have been challenging for a language model to write a corona from start to finish without any notes, just as it would be for a human. It's much more practical to plan ahead — in particular, to write the master sonnet alone, trying to choose lines that could open and close their own chapters.
My initial prompts explained the overall goal, the structure I was aiming for, and spelled out in loose terms a workflow we could adopt. I asked Claude to begin with a theme and mood, then turn that into a rough story, and only then to draft the master sonnet. Claude came up with a list of potential themes, and the one I liked best was the elderly artist accepting that the arc of their story was coming to a close; the mood was philosophical, introspective. Once the master sonnet had reached a first draft, Claude spun up 14 simultaneous agents to independently write the other 14 sonnets. Things were moving along!
The first draft had many issues. The vast majority of constraints were correct: rhymes, meter, and matching lines. But a fraction needed correction. Many lines did not flow musically. The descriptions tended to be literal. The story meandered across the 15 sonnets. A hint would be dropped of a critical thread — a missing daughter, maybe? A wife who left, or died? But those hints felt clumsy and lacked closure. They landed more as aberrations than the contours of a deliberate journey. The time of day would jump; other details were inconsistent.
Over a series of 11 versions, I asked for many iterative changes to fix the less subtle of the issues. An easy change to request, for example, was to avoid the over-use of a single word, such as "light," which still appears 7 times in the finished corona — down from 14 in the first draft. Claude was capable of less obvious changes as well. It seemed to understand meter to a decent extent, and I could effectively ask that some lines be re-written for better sound, or to better connect the meaning of the surrounding narrative. I chose to edit in a conceptual top-down manner, addressing deeper issues like a confusing story, or rambling sections before focusing on finer points such as individual word choices. I spent the most time on the master sonnet, which impacts all of the others, and then smoothed out the biggest wrinkles that remained in the rest.
I spent a total of $43.57 to make the corona. I could have spent a small fraction of this if I had stopped at the first draft. The API time spent — not wall-clock time, but time in which the API was busy — totalled 168 minutes.
The Result
The final product is here.
I had informally asked some earlier LLMs to write shorter bits of poetry. I haven't saved those. But in the light of comparison, and being impressed that AI can do this at all, there are many things working well. I'll focus on the final sonnet, which includes some suggestions from me — but I'll try to discuss what Claude wrote of its own initiative:
The evening comes; I paint by fading light. I learned to see before I learned the trade. These hands have traced the arc of what I've made. The canvas waits, stretched taut and winter-white. This room has been my eye, my second sight. There was a summer when the colors stayed. And yet true strokes are few; most only fade. The shades have pooled and spill into the night. In silence have I worked, and work here still. What could be shared, I shared, and let it go. And even things undone have left their mark. What's left is for the future, good and ill. What lasts of all I've done I cannot know. A hue, a spark, a debt called due by dark.
With some generosity, these things work:
- As a baseline, all of the structural elements are correct.
- The theme and mood, if uninventive, are at least conducive to the form and a sense of something profound.
- There are clear elements of symbolism: summer and winter, light and dark, things unfinished, or the potential of an empty canvas; and questions of legacy, ownership, humility and hope in the things we see, the things undone, the things left, a spark.
- Some lines, to my ear, read nicely. "I learned to see before I learned the trade" begins to fill out a life story. "In silence have I worked, and work here still;" a solid meter, and a pleasing rhythmic beat.
- There is some intentional repetition that adds to the rhythm and carries us forward: "learned", "work", and "left." In conversations, Claude clearly understood what's good about hitting these notes with purpose versus the blemishes of a limited lexical palette.
While not clear from the master sonnet alone, I also give credit to Claude for directing the course of a story throughout the other sonnets, chapters that led naturally to each other, and that flowed within themselves. Consider sonnet 2:
I learned to see before I learned the trade. The dust motes swirling in a shaft of sun were all the world I needed, every one a planet drifting where the beam had strayed. I watched the darkness gathering, unafraid, beneath the chairs where deeper rivers run, and knew the world was never truly done— each hour the colors built a new parade. What gifts are these that come before we ask? The eye that drinks the world without a name, the hand that reaches, small, toward light displayed— the child who stares, then mirrors with his task, whose breathing grows to tend the fanning flame. These hands have traced the arc of what I've made.
Unlike the master sonnet, I had little to do with this one. The astronomical metaphor is all Claude, and I like it. This chapter, this sonnet, works not only within itself, but within the corona as a hint about the painter's origin — how he (she?) thought and learned early in life.
Having said all that, I would not be satisfied if I had written this myself, nor would I feel overwhelmed if I read this corona, thinking it was by a human. I honestly don't know if I'm being harsh or naive. Though I love to read, literary criticism is far from my wheelhouse. Nonetheless, I'll share why I think the corona is interesting but not great.
The biggest problem is that the poem isn't saying anything new. For me to connect with a work, I want to feel or learn. I don't have to agree, but I want to care. The strongest feeling I have when reading this is akin to: "it's neat that AI can do this."
While I'm happy with the technical execution, it's not subtle. While I like the metaphors, they don't speak to me as something achieved or inspired. I want to hear a song well sung, see a picture well painted, feel an emotion that plumbs depths. I want layers and nuance and a world adorned with the unexpected idiosyncrasies that make it real.
But that's not the right note to end on. I admit that I'm biased in the world of AI because I've studied how language models learn; I've been paying attention to how they think and report their inner experiences. Most readers won't feel like parents to these models. At the same time, Claude has learned from us as a collective species. It has inherited so much of how we think. Not just a few of us, but many of us. Like it or not, this intelligence is borne of ourselves.
I'm not deaf to the intense dislike many have toward the nexus of AI and creativity. There's a justified disillusionment with low-quality work — AI slop. There's anger and disgust at individual or well-known artistic styles being copied. And a sense that AI is fundamentally artificial, lacking in genuine originality. This little note isn't the place for that debate. I know it's there, and I'm not asking about what's right, or what we should do.
I'm asking instead what's possible, and getting a sense for what might be next. In that light, this corona is a first step — a step that, at least for me, bestows wonder and curiosity at what the future holds.
You can also read Claude's analysis of its work here.