At LavaCon 2026, a poetry workshop set the theme for the whole conference. Talks on AI agents, taxonomies, and canon kept coming back to the skill poets practice, choosing the specific word or object, and that practice is how content professionals build the judgment Aristotle called phronesis.
The first thing I did at LavaCon this year was write a poem.
If you haven’t been, LavaCon is an annual conference for content strategists, technical writers, and content operations folks, run by Jack Molisani. This year it was in Charlotte, and almost every session was about what AI is doing to our work.
But that’s not where it started for me.
Carol Hattrup runs a free pre-conference poetry workshop every year called “They Already Sent a Poet, and It’s You!” Her premise is that content people are already poets. We just don’t practice.
I didn’t need convincing.
My path into rhetoric started in a class on the rhetoric of surrealism, and my favorite activity was the exquisite corpse. One person writes a line of poetry, folds the paper over, and passes it to the next person, who adds a line without seeing what came before. Nobody sees the whole poem until the end, and it always surprised us. That game is still how I think about writing with AI.
But it had been a long time since I’d sat in an actual poetry workshop, pen in hand ... and I didn’t expect it to feel so much like a design thinking session ... something I have been doing a lot with students and professionals.
Carol opened with a group poem built from her version of Jim Simmerman’s “Twenty Little Poetry Projects,” a list of odd instructions:
Use a name.
Say something preposterous.
Have the narrator do something impossible.
Contradict the line before.
We called out lines one at a time, and someone wrote them down.
It worked a lot like the exquisite corpse. Each instruction kept any one person from steering the poem, and the surprises came from the constraints, not from anyone being brilliant.
Poets know word choice is important ... now with AI, word choice is infrastructure.
Design thinking calls this divergence. You set rules that force you past the first obvious idea and hold off judging until you have more than you need.
Then the workshop moved toward convergence:
Go out and find one ordinary object.
Draft without stopping, circling words to fix later.
Edit with Carol as the editor.
Read the poem aloud to the group.
Observe, prototype, iterate, test. The reading did one more thing. Carol said she never meant the workshop to be team-building, but it becomes that every year, because sharing a draft that rough builds trust fast.
And we walked away with something none of us could have written alone, which is what a good design thinking session produces too. Our group poem has a man named Tingle Lips running through the hotel halls in a cape, even though he has no legs, and llamas eating warm chocolate chip cookies. Together we made a new kind of sense out of it. That’s the point of both poetry and design thinking.
So what does poetry have to do with content strategy?
Well, poets make great content people.
Poets already know information types
During the workshop, Carol showed us how content people already know two forms of poetry.
The first is the list poem.
It’s a catalog, often with each line starting the same way, and it builds meaning through repetition. Carol traced it back to Homer’s catalog of ships.
The second is the how-to poem, which gives the reader instructions. The best ones also say what not to do, and they bend the steps toward something no manual would cover.
I heard those two forms and thought of information types.
A list poem is reference, and a how-to poem is a task. Technical writers have spent decades learning to keep those two apart. Poets have been using them as shapes for art much longer.
Recently I separated genre from information type in my own knowledge base. Genre is what kind of document you’re holding. Information type is what each chunk inside it does for the reader.
Carol’s forms suggest these labels go deeper than filing. They show up wherever people organize language, in a reference manual or in a poem.
I’ve been arguing that the five information types reflect how we sort knowledge in our heads. A list poem and a how-to poem are some of the oldest evidence for that I’ve come across ... and I want to dig into this more.
Good Poets (and content professionals) Go Small
The key to good poetry is being concrete. That’s why Carol sent us out to find a “thing” during the work shop. My ordinary object was a Dum Dum lollipop. A mystery flavor, still in the wrapper.
For example, in my poem, I talked about throwing documentation into a “rumbling wood chipper,” instead of talking abstractly about AI.
Writing specific things is how we learn to judge them.
What I noticed in the workshop was how fast the poem went flat whenever I drifted into big ideas about surprise or writing. It only came alive when I stayed with the lollipop itself, with its wrapper and its flavor I couldn’t guess.
That turned out to be the lesson of the whole conference. Content works best with AI when you are concrete and specific.
On Monday, Manny Silva showed how AI agents can repair broken documentation overnight. His rule for the repair agent was almost a poetry prompt.
Fix the three pages.
Swap the screenshot.
Stop.
Anyone who has worked with an eager agent knows what happens without that rule. You ask for one change, and it helpfully rewrites five more.
Stefan Gentz made the same point from the reader’s side. A technician asks an AI assistant, “Can I restart this device now?” A good answer depends on which device, which software version, and what steps came first. Drop those conditions and the assistant gives a confident answer that might be wrong.
Sarah Dugan, a documentation lead at GitBook, took this lesson back almost two centuries.
In the 1840s, the mathematician Ada Lovelace translated an Italian engineer’s paper on Charles Babbage’s Analytical Engine, a steam-powered calculating machine that was never built. She added notes of her own that ran three times longer than the original.
In the last of them, Note G, she wrote out step by step how it would calculate one specific sequence of numbers (the Bernoulli numbers). It’s often called the first computer program, and it started with one concrete example.
Poems and agents fail the same way ... they are vague when they should be concrete.
Words are infrastructure
Poets know word choice is important ... now with AI, word choice is infrastructure.
For example, Christi Guzik runs content operations at Rocket, the mortgage company. Three years ago, she went looking for who defined the names in Rocket’s analytics data. A developer told her, “Oh, I just make it up.”
So she built a language. Every event name is a past-tense verb followed by a noun. Everything is lowercase with underscores. There’s a list of words to use and words to avoid, so you click, you don’t tap.
She jokes that people sign up for taxonomy training and find themselves in an English class.
The payoff came with AI. Rocket’s analytics tool now answers questions in plain English, but only for teams that adopted the standard. Teams that held out found their data missing.
Rashmi Ramaswamy saw the same thing from the documentation side at Guidewire. Their AI assistant scored zero on helpfulness for questions about one product. Users kept typing “IG.” The writers had been told to always spell out “Integration Gateway,” so the assistant never connected the two.
Adding a synonym list took helpfulness from 0% to 90%. One word made that whole difference.
I found the same problem in my own knowledge base this summer. An audit turned up 1,988 unique tags, and at least 45 pairs named the same idea split only by capitalization, like rag and RAG or ai and AI.
Then there were the synonyms. Tech-writing, technical-writing, and technical-communication all meant one field. Every search on those topics was quietly returning half of what I had.
So I built a tag thesaurus, with one preferred label for each idea and the synonyms pointing to it. It’s slow, unglamorous work, but now key to making our content work at scale.
Every entry was a small choice between near-synonyms, the same kind of choice a poem asks for line by line.
At ConVEx last year, the conversation was about purpose. At CAKE last fall, it was structure. At CAKE this fall, it was content engineering. At LavaCon, it was language ... the plain, careful choice of what to call things.
Who decides what’s true
The last lesson came from Plato, by way of Keith Boyd, a senior director at Microsoft.
He opened Tuesday’s keynote with a story from the Phaedrus. The god Theuth offers writing to an Egyptian king as a cure for forgetting. The king turns it down. Writing, he says, will make people seem wise without being wise.
We only know that warning because Plato wrote it down. And Boyd’s point was that we’re in the same spot again. Fluent language is everywhere now, so the scarce thing is someone who can tell what’s actually true and what matters.
The talks that followed kept asking who that someone is.
Rob Hanna of Precision Content answered with a test. Any AI answer should trace back to an approved source, in the right version, as of the right date. If it can’t, it isn’t canon. But a test only works if someone decides which document counts as canon in the first place.
Jason Kaufman showed what happens when nobody makes that call. A credit union asked its new chatbot who the CEO was, and it named the CEO from ten years ago. The AI was faithfully repeating something that was way outdated.
And Sarah O’Keefe pointed out that even a good answer won’t stay put. Readers can now ask any AI to shorten, translate, or rewrite whatever we publish. We used to decide which versions of our content existed. Now readers get the last word.
So the judgment can’t be built once and handed off.
Aristotle called it phronesis, the practical wisdom to make a sound call in a particular situation where no rule quite applies. After CAKE last month, I asked whether it can be taught.
LavaCon gave me a better answer than I expected. Poets practice it all the time.
You don’t learn phronesis by memorizing rules about good writing. You learn it by choosing between unwrapped and open, or between the lollipop and the idea of surprise, again and again, until you can feel which one the poem needs.
Hanna’s canon test depends on that same judgment, and so do Guzik’s preferred labels and Gentz’s conditions.
Writing specific things is how we learn to judge them.
Still wrapped
I read my workshop poem at LavaCon’s storytelling night on Monday, in a dim room with no microphone, between a story about Alan Greenspan and one about a crown from a Bahamian parade.
It ends with the lollipop, still in my pocket, still wrapped. I never did find out what flavor it was.
That felt right for the week. Every talk I heard was about making things specific enough to trust, whether a page, a label, a condition, or a single calculation. But the reason any of it matters is because someone decided what’s worth saying in the first place.
Poets have been practicing that for a long time. Content people have too, and we’re about to need it more.
If you want to read the poem itself, I posted it on my poetry Substack: The First Thing I Did at LavaCon Was Write a Poem.