By Jim Gill, Head of Content & Storytelling, Blickstein Group
If you want to use AI well, you need to think like (or hire) a humanities professor to build your prompts. I say that without the winky face emoji or the ever-present “jk” I add when I’m being snarky in a text. Using AI well is not about writing better tactical prompts but instead about being able to articulate, in writing, how you—the human user—actually think. Because, despite near-Turing-test appearances and the claims of the tinfoil-hat crowd, AI can’t think for itself.
Before I became a content strategist in the legal technology industry, I taught writing and literature at the college level for 16 years. When it comes to developing AI prompts, this might be an unfair advantage since I also have a decades-long archive of lecture notes, handouts, and writing exercises. This is a physical, accumulated record of trying to teach other people how to read carefully, how to make an argument, how to spot logical fallacies, how to analyze texts and sources, and how to know when a sentence is doing what it claims to be doing. So when I started using Generative AI in my content strategy work, I decided to put my professor’s cap back on. And I have to say, Claude has been an excellent student.
Most people talking about prompt engineering are really hunting for quick phrasings that produce better output. But there’s no easy button. Not even with AI.
The reason most AI-assisted writing is bad is that the people using it have never articulated for themselves how they think, much less have it written down. They may have a document from their company that provides a bit of information about brand voice. They might even have a list of forbidden words or font styles or a policy on the Oxford comma. So when they ask the AI to write something, they are asking it to imitate surface-level outputs, and AI is very good at that. But the old saying is still true: garbage in, garbage out. And while it may not appear to be garbage, it’s easy to see AI-assisted writing as homogenized fairly quickly. Because the prompts don’t ask for anything more.
So when I started creating my “house style guide,” I decided to treat the AI (Claude in my case) as I would one of my students, many of whom were excellent at getting good grades, but few of them had learned how to develop layered, complex thoughts and then communicate those thoughts clearly and effectively. My style guide soon became an architecture for how I reason and a framework based on all the elements that communicate that reasoning. And that document is ongoing, not static.
Outside of my work as a content strategist, I’m a writer and have published two books of fiction, as well as several creative non-fiction pieces published in literary journals (none of which were created with the help of AI, I might add). When writing my latest novel, I hammered out the entire first draft on two mid-century portable typewriters. Forward motion only, no backspacing, keys that stick, ribbons that slip. In other words, nothing is perfect. The next stage is revising by hand, reading carefully, marking the manuscript up with a pen. Then I type those changes into the computer, and the editing and polishing work begins. That’s the way I’ve written more or less for the past thirty years. It’s slow and measured. Exactly the opposite of how AI produces things.
Now that I work with organizations using AI in their workflows, and some that are building AI tools, I’ve started thinking that part of the problem is speed. AI produces polished outputs in minutes, sometimes seconds. So I started looking for ways to get AI to slow down and go through the process with me.
The opening section in my “house style guide” is about the composition process, the stages that transform thought into a piece of writing. Then it includes my requirements for the writing itself: a catalog of opening gambits, an Aristotelian formula for thesis development adapted for modern English, the use of proof and evidence, the development of meaningful conclusions. Further down are specific strategies for sentence-level strength.
The composition process, the part where you teach the AI how you think, is half of the work. The macro arc of how a piece develops. What the work is supposed to look like before it goes anywhere near a sentence. The downstream half is execution.
I’ve taught Claude to deliver what I call “typewriter drafts.” Intentionally unpolished first drafts whose only job is to get the thinking on the page. I have Claude deliver the outputs in Courier font in files I can edit within an AI chat thread, the way I’d mark up a manuscript by hand, so there’s back-and-forth with Claude, refining things before any of it lands in a doc.
The style guide also builds in QA layers that Claude runs to check outputs against the standard. There are sections dedicated explicitly to working with AI. One catalogs the telltale signs that give away generated prose. Another is a set of named prompt blocks that keep Claude from fabricating sources, statistics, or analyses it does not have. Another is a list of jargon, often mimicked by AI but long a symptom of bland corporate copy before AI even existed. There are references to specific parts of my favorite style guides—Strunk & White’s The Elements of Style, Gardner’s Art of Fiction—for the sentence-level work. And I can run a QA pass from each of them: a jargon pass, an “AI tells” pass, an AI failure modes pass, a Strunk & White pass. I’ve also set up templates for client-facing work and writer-facing work, each with a different tone and length.
Rather than trust the editing to the AI, I will do the editing that I’ve always been good at, and then ask the AI to check my work. Am I missing anything in the original project guidelines? Does this meet the client’s objectives? Does this match my own style guide, or have I gone off course? Are there any sources or quotes that should be confirmed by human review?
The usual approach to prompt engineering treats the prompt as a clever instruction, a phrasing trick, a way to hack model behavior. I’ve tried to treat the prompt as the transmission of expertise from someone who has it to a system that does not. This is not merely a matter of polishing passable prose. It’s a densely constructed (currently 25 pages) document that I begin every project with as a baseline of what’s expected before I ever give the parameters of the actual task at hand. I still write task-specific prompts for each project—what to write, for whom, in what format—but they can stay short and focused because the baseline has already explained how I think, what good looks like, and what to watch out for.
The lecture notes, the handouts, the writing exercises, the marginal comments on literally thousands of student papers—the accumulated record of articulating things I knew, for students who did not yet know them—turned out to be the raw material the AI needed. AI needs a teacher, not just a prompt. To use AI well, the user has to take their own thinking and put it into language that a system with no thinking of its own can use. Which is what humanities professors have been doing all along.