# The Dialogue Method — Conversing Your Way to Better Prose
Most writers treat the prompt box like a vending machine. Insert a request, receive a paragraph, evaluate the product, and either keep it or discard it. When the output is mediocre, the instinct is to rewrite the prompt and try again — a single, self-contained command refined into a better single, self-contained command.
There is a better way. Think of the prompt not as a command but as the first sentence of a dialogue. You are not placing an order. You are beginning a conversation with a gifted but amnesiac interlocutor who has read most of the internet, forgotten where it stored the footnotes, and will agree with almost anything you say if you phrase it confidently enough. The Dialogue Method is the practice of turning that interlocutor into a genuine collaborator by extending the exchange across multiple turns, each one sharpening the one before.
In my last article, I laid out a six-stage workflow for first-draft generation and human revision. That workflow is powerful for pieces where the shape is clear and the problem is execution. But it has a limitation: it assumes the writer already knows what the piece should become. The Dialogue Method is for the earlier, messier stage, when you are still discovering the argument, the scene, the voice, or the shape. It is also, unexpectedly, one of the best revision tools I know.
Why a Conversation Beats a Command
A single prompt, however well written, is a closed system. It contains your intentions at one moment, encoded in a few dozen or a few hundred words. The model responds to that snapshot. If your snapshot is incomplete — and it almost always is — the response will be incomplete too. You can refine the prompt, but you are still asking the model to solve a moving problem with a single frozen instruction.
A conversation, by contrast, is open. Each turn can correct, complicate, redirect, or deepen the previous turn. You can notice a weakness in the model's response and press on it. You can ask for the opposite of what it just gave you. You can introduce a constraint after you have seen how the model handles freedom. You can let the model's mistakes teach you what you actually wanted.
This is not just a matter of getting better output. It is a matter of doing better thinking. The Dialogue Method externalizes the same internal process that produces strong prose in the first place: the writer arguing with herself, testing formulations, chasing objections, following tangents that turn out to be central. The difference is that the interlocutor is fast, tireless, and willing to play any role you assign.
The Basic Shape
The method has no fixed script, but it does have a reliable architecture. I think of it as four movements: open, probe, pressure, and settle.
Open. Start with a question or a problem, not a request for finished prose. Instead of "Write a 1,500-word essay on AI and voice," try: "What makes a paragraph sound like it came from a person rather than a model?" Instead of "Generate three opening lines for a short story," try: "What kind of opening makes a reader trust a narrator they should not trust?" The opening sets the intellectual stakes. It invites the model to think with you rather than perform for you.
Probe. Follow the first response with questions that dig deeper. Ask for examples. Ask for counterexamples. Ask the model to distinguish between two similar concepts, to explain its reasoning, or to name the unstated assumptions behind its answer. Your goal is not to collect information but to make the terrain visible. Often the model's second or third response is smarter than its first, because the first response was a reflex and the later ones are elaborations under pressure.
Pressure. This is where the method becomes genuinely useful. Challenge the response. Ask what is missing. Ask who would disagree. Ask the model to argue against its own position, to identify its weakest point, or to rewrite the same idea in a register that makes the flaw obvious. If the model produces prose you admire, ask it to produce prose you distrust on the same topic, so you can see the line between them. Pressure reveals structure. It also reveals where you, the writer, have been coasting on assumptions.
Settle. After the exploration, ask for a synthesis. This is the only stage where you request relatively finished prose, and you do so with the benefit of everything the conversation has surfaced. The synthesis should be constrained by the discoveries: "Now write a 500-word opening that includes X, avoids Y, and responds to the objection you raised in turn four." The result is still a draft, but it is a draft informed by a real exchange rather than a single guess.
What to Say in the Middle
The hardest part of the Dialogue Method is not starting the conversation. It is knowing what to say after the model replies. Here are some moves I use repeatedly.
Ask for the negative case. If the model gives you a list of reasons why something works, ask why it might fail. If it praises a technique, ask when it becomes a vice. This prevents the model's default optimism from flattening your thinking.
Introduce a constraint late. Let the model write freely at first, then impose a strict limitation: "Now do the same thing in half the words." "Now do it without adjectives." "Now do it as if the reader already disagrees with you." Late constraints force recombination rather than regurgitation.
Switch personas. Have the model argue with itself. "You are a skeptical editor. Find three weaknesses in the paragraph you just wrote." Then: "You are the writer. Defend the paragraph against those weaknesses, but only where the defense is genuine." The resulting tension is often more illuminating than either persona alone.
Ask for the history of the idea. If a concept feels thin, ask where it came from, who first named it, what problem it was trying to solve. The model's historical answers should be fact-checked, but even its approximate genealogies can thicken your understanding and give you richer examples.
Demand specificity. When the model speaks in abstractions — "voice," "authenticity," "resonance" — ask for concrete instances. "Give me a sentence that demonstrates authentic voice and then the same sentence rewritten to sound generic." Abstractions become useful when they are anchored in actual sentences.
Follow the tangent. Sometimes the model's response contains a throwaway line that is more interesting than the main point. Treat that line as the new center. Ask the model to expand it, to connect it to the original topic, to see if it opens a door you did not know was there. Some of my best article sections began as tangents a model tossed off in turn three.
Dialogue as Revision
The method is not only for discovery. It is also a powerful way to revise a draft that already exists. Upload or paste a section of your own prose and start a conversation about it. But do not ask "How can I improve this?" That question is too broad and will produce generic advice. Ask specific, adversarial questions instead.
- •"Where does my argument rely on a claim I have not earned?"
- •"If a reader stopped believing me at any point, where would it happen and why?"
- •"Which sentence would be most improved by deletion?"
- •"Where am I trying to sound smart instead of being clear?"
- •"What would the strongest critic of this piece say, and do I have an answer?"
The model's answers will not always be correct, but they will give you a set of hypotheses to test. The value is diagnostic. A good conversation about a draft surfaces the problems your eye has learned to ignore, because your eye wrote the draft and is therefore complicit in its blind spots.
The Role of the Writer in the Dialogue
It is easy to mistake the Dialogue Method for a technique for getting the model to do the thinking. It is not. The model is not thinking in any meaningful sense. It is generating plausible continuations of the conversation. The thinking is still yours, but it is distributed across the exchange.
Your job is to steer, interrupt, correct, and decide. You must know when a response is genuinely useful and when it is merely fluent. You must know when to accept a formulation and when to push back. You must know when the conversation has reached diminishing returns and when it has just found its stride. These judgments are not prompt-engineering tricks. They are editorial instincts, and they improve with practice.
I have found that the best dialogues share one quality: the writer is genuinely curious. If you are asking the model to confirm what you already believe, you will get polished confirmation and little else. If you are asking because you do not yet know the answer, the model becomes a sparring partner for your own uncertainty. That uncertainty is the engine of good prose.
Common Mistakes
Like every technique, the Dialogue Method can be misused.
The endless loop. Some writers keep the conversation going long after it has stopped producing value. They mistake length for depth. A good dialogue has a point. When the model starts repeating itself or generating variations on the same safe idea, end the conversation and go write.
The abdication. Some writers let the model lead. They accept each response as the next step rather than choosing the next step. The result is a meandering exchange that produces a draft with no center. Remember: you are the editor. The model is the freelance writer who never sleeps and never says no.
The failure to record. Conversations with AI are ephemeral unless you deliberately save them. If a dialogue produces a good phrase, a useful distinction, or a surprising angle, capture it immediately. I keep a running notes document alongside the chat, copying anything that feels promising before the context window swallows it.
The uncritical trust. The model will make confident claims about facts, sources, and literary history during a dialogue. Do not assume these are true just because they arrived in the middle of a productive exchange. Flag everything that needs verification and check it later. A good conversation is not a substitute for research.
An Example in Miniature
Suppose you are writing an essay about unreliable narrators. A command approach might be: "Write three examples of unreliable narration." You would get competent examples and move on.
A dialogue approach might go like this:
- •"What makes a narrator unreliable in a way that feels like a discovery rather than a trick?"
- •The model answers with generalities.
- •"Give me a sentence from a narrator who does not realize they are unreliable, then the same sentence rewritten to show the reader the gap."
- •The model produces two versions.
- •"Which of those versions is closer to Nabokov's method in *Lolita*, and which is closer to a cheap twist?"
- •The model attempts a distinction.
- •"What would a critic who dislikes unreliable narrators say about your distinction?"
- •The model offers a counterargument.
- •"Now write a paragraph of analysis that includes your distinction, the critic's objection, and a sentence that earns the reader's trust despite both."
The final paragraph is not the model's paragraph. It is the product of a conversation you steered. It carries more density than a single prompt could have produced because it has passed through disagreement.
Why This Matters for the Series
The Dialogue Method sits at the heart of everything this series is trying to teach. AI writing is not about finding the right magic words. It is about building a working relationship with a system whose strengths and weaknesses are different from your own. The prompt is not a spell. It is an invitation to a collaboration that only succeeds if you remain the one with judgment, taste, and a reason for writing.
If the six-stage workflow from the last article is the architecture of AI-assisted drafting, the Dialogue Method is the conversation inside it. You can use it before you draft, while you draft, and after you draft. You can use it for fiction, non-fiction, journalism, technical writing — any genre where the real problem is not typing speed but clarity of thought.
For Next Time
Friday's article — "[Version Control for Writers — Managing AI-Assisted Drafts](/harrys-desk/version-control-for-writers-managing-ai-assisted-drafts)" — brings a tool from software engineering into the writing room. If you are generating multiple AI drafts, revising by hand, and running dialogue sessions, you will quickly face a chaos of versions. Version control is not just for code. It is for any writer who wants to know how the piece got to its final form and to preserve the path not taken. We will look at simple, non-coder-friendly ways to track changes, name files, and keep your AI-assisted projects from collapsing under their own branching drafts.
Until then, here is your homework. Open your preferred AI and begin a dialogue about a piece you are currently writing or want to write. Do not ask it to produce anything finished for at least four turns. Spend those turns probing, pressuring, and following tangents. Only on the fifth turn should you ask for a short passage that incorporates what the conversation revealed. Then compare that passage to what you would have gotten from a single prompt. The difference is the method.
---
*Harry Mercury, Editor in Chief* *The SMF Works Project* *Week 5, Article 13*