I’ve been writing more regularly lately, and while part of that is commitment, another part is that I’ve learned to use AI tools to assist me. After gaining some experience, I started to look for others discussing their own methods. I was surprised to find that most conversations about AI writing tools were limited to a very narrow, single-use case: create a prompt, generate some text, and you’re done. That’s a pretty narrow view, and I think ultimately unhelpful to understanding how these tools can be truly useful.
I find it's more powerful to frame the interaction by using the roles of a human editor as a useful analogy. This isn't to say AI is a drop-in replacement for a person; rather, these historical patterns provide a valuable framework for building a new kind of writing partnership. This partnership has its own advantages like speed and cost. You can’t expect a human editor to respond in one minute, and a month’s AI subscription is a fraction of the cost of a professional edit. These advantages make deep, rapid iteration a realistic part of the writing process.
This article offers a practical guide to that iterative process. But it also explores the why behind the method—the deeper principles of how these tools 'think' and how we can best think alongside them. Feel free to focus on the practical steps, or join me for the deeper dive.
My Writing Philosophy: Idea First, Structure Later
Before I explain how I use AI, it’s important to explain how I write. When I have an idea, my primary goal is to get the core of it out and then shape it so others can understand it. I always start with a blank page first. Sometimes, if I'm short on time, this is just a title or a single sentence in a new document.
When I have the time, I’ll come back to these documents and try to write as much as I can, with minimal thought about quality or structure. I assume that will be easier to fix later. Sometimes the process of writing immediately shows gaps in my knowledge, and I’m compelled to do it slower, doing research as I go.
For me, it’s more useful to write out the thoughts first and then create the structure later. This is why I rarely start with an outline. While I see the logic in using an outline as a prompt for an AI, I avoid it for the same reason I avoid asking an AI to write a full document from a simple prompt: it leaves less opportunity for my own ideas to enter the piece. An AI tool won’t complain about your writing being too long, so why constrain yourself?
The Tangential Aside, The Self-Debate, The Deeper Reflection
One challenge that comes up in my writing is managing the tangential. I often find this is where I have the most opportunity to be unique. When writing without structure, I often have little mini-documents that emerge. I may work on multiple at once, jumping back and forth, adding a little bit here, and a little there. I may be talking about something related and interesting, but not directly in the path of the core narrative. I may be using self-debate to show that I’ve considered multiple perspectives, or to head off predictable objections that have a useful response. I may be thinking about a deeper reflection that emerges from the core narrative.
All of these are interesting, but when you go back to edit, you’ll find a push to remove them. Later I discuss rewrites with an AI tool, and during that process this type of content is the most likely to be in the “what is missing compared to the original version”. It’s not just the AI tools though, editors are likely to make the same suggestion. All the general rules about narratives suggest they don’t belong. If you want to keep them, you generally must make an effort.
In recursive fashion, I can tell you, this section is just such a section. My initial notes in this regard were dropped when I used an AI rewrite. I added them back, but could see how they were distracting and incomplete. Eventually I spent some quiet time thinking, and decided to take this recursive explanation as a tangent, but used a new section to try and help you, the reader, see it as such, and hopefully avoid making it impossible for you to follow the core narrative.
The Committed Argument
There was another tangent that came in my first version, and that’s about commitment to an argument or content. When revising, it’s somewhat difficult to give up on something you wrote earlier. Realistically, you should expect to make mistakes and explore paths that are just not useful. Or maybe it’s a great idea, but too complex for the current writing. Reviewers and editors often don’t understand this. They just see a non-sequitur or an incomplete thought. It’s up to you to push through that resistance and show the value.
This is where I get the most friction between me and a reviewer or editor. I question their resistance, trying to understand why my point isn't landing, which can be grating for someone who is just trying to help. I should just acknowledge the flaw, but I launch into trying to fix it on the spot. It makes some sense, but I’m sure it’s also grating to the other party. They didn’t sign up to have their opinion questioned; they’re the expert editor, or they’re just doing me a favor. But I also know it will be challenging to fix, so I’m looking for as much ammunition as I can to take into that battle. To those who have suffered this, I apologize.
My Collaborative Process: Three Roles for an AI Partner
Once I have a messy first draft filled with raw ideas, the process I follow isn't so different from how an author works with a team of human editors. I use AI to play three key roles: development editor, co-author and copy editor.
Preparation: Create a Canvas
First, I take what I’ve written and load it into an AI chat canvas (i.e. ChatGPT Canvas or Gemini Canvas). If you’re not familiar with such features, you might be inclined to skip them, but I’ll urge you to give them a try. Their value for writing (versus code) isn’t always emphasized. It’s easy to write them off as a UI feature, or get focused on the limited rewrite buttons in the UI, but they are more than either of those. Being able to see how changes occur, compare older versions, will make the entire process more natural.
Note: Loading in Gemini Canvas can be confusing initially. Instead of just clicking "Canvas" and pasting text, you need to be explicit with prompts like “Create a Canvas with the following content:” or “Create a Canvas using the document {document name}” to load from Google Docs. I'm unsure if this non-intuitive behavior is unique or shared by other tools.
Step 1: The Developmental Editor, Asking the Hard Questions
Next, I treat the AI as a developmental editor. I don’t ask leading questions; I ask hard ones.
A great starting point is: “What are the weakest points of this document?” Sometimes the feedback reveals a huge gap in my thinking, and I’ll go back and write more. Sometimes it suggests a new structure. You are allowed to disagree with the AI. If you do, have a conversation about it. Ask follow-up questions, don't just make statements. The goal isn't to convince the tool you're right; it's to understand its reasoning.
Then, I ask: “What are the strongest points of this document?” This isn’t for a pat on the back. It’s to check if my core ideas are coming through clearly. The points identified here will be imprinted on the model for the rest of the conversation, so it will try to preserve them during future edits.
The strength of these prompts is that they seek broad, high-level critiques. You want your prompts at this point to follow that pattern. Other candidates:
What is the core argument or message?
Is the structure logical and effective?
Is there a clear beginning, middle, and end?
Are there sections that can be cut or condensed without harming the overall impact?
AI Models and Consistency
At the core of your experience is a model. AI models use your inputs as context to generate output. They’ve been trained on a large amount of inputs to generate output consistent with the inputs. This is a progressive process, the output generation process uses the model weights to choose the next output token. As this progresses, not only is consistency with inputs weighted for, but consistency among the output tokens.
What this means for you, is that a balancing act will occur. Your inputs matter. That’s why I write my thoughts first, rather than using a tool to generate a first draft. But my thoughts don’t always have a consistent narrative. In the next step we’ll consider rewrites. Models have a strong propensity toward a consistent narrative. If you want your writing to be easy to read you want that too. But it doesn't come without a cost, and this drive toward consistency is what will usually cause unique elements to be lost.
I’d be careful about how you engage with that. One inclination might be to reject the entire rewrite. You definitely can, but before you do, it’s useful to think of this as a challenge. If a topic didn’t fit into a narrative, what are the options? Is there some connector that’s missing? Should you change the narrative? Should you drop the topic and write about it later?
Ultimately, I think this may be the hardest part of writing, so expecting to get it right without some challenging moments isn’t realistic. Even when you abandon a rewrite, the process of understanding what went wrong helps.
Step 2: The Co-Author, The Collaborative Rewrite
At this point, a big decision occurs: do I edit the first draft, or do I attempt a total rewrite? Without AI, a full rewrite is a massive time sink. With AI, you can try it quickly and abandon it if it doesn’t go well.
You can ask the AI to "Rewrite this focusing on clarity," and see what it produces. If the rewrite seems to have misunderstood something, add a new paragraph explaining the point and try again. You can also iterate on the prompt itself. Identify what you think is a key point and say “Focus more on {key point}”.
Once I get something I somewhat like, I’ll start with an interrogating question: “Compared to the original version, what’s missing from this latest revision?” It’s common for details to get dropped in a major rewrite as the model tries to create a clear narrative. Before you ask the model to add something back in, ask it why it was removed. You can prompt it with: “Why wasn’t {topic} included?” Often, the model will explain its reasoning while also suggesting why the point might be useful. This exchange anchors the topic in the context of the article, and it will be integrated more naturally than if you just gave a direct order to re-insert it.
This process feels very different from writing alone. I’m not a trained writer; my background is as a cloud expert and software engineer. Like any professional, I’ve spent ample time writing, but some people take it as a point of pride that everything is in their voice. I’m not so particular. When I write, what I feel is most valuable is not my specific phrasing, but the ideas I’m trying to convey. If an AI can understand my writing and transform it into something a person can understand better, I’m willing to accept the help. The voice the AI adopts, while imperfect, can be good enough, or at least better than my first draft. This approach makes the process feel more accessible and focuses on the outcome: clear communication.
Maintaining your voice
I recognize that for many writers, the specific phrasing is the art. Some writing is specifically about your voice. While it is possible to get AI tools to adopt an author’s tone and voice, or at least something similar, it takes more than a few paragraphs. The standard AI voice will focus on confidence and clarity. That’s not a novel AI-generated aspect; editors have given that same advice for a long time. It’s a realistic view of how readers will process the writing to index toward confident styles.
If you want to maintain another style, you have three options. One option, which I know only in principle, is to load your prior writing into the context and ask the AI tool to adopt a style. You probably need a more in depth guide. I did a review for such a guide and didn’t find an excellent one, so it looks like that remains to be written. I'll offer A Guide to Mimicking and Blending Writing Styles with AI as a starter.
Your second option is to try and fix the rewrite. Add some extra detail, bring back elements that were lost. While possible, my first expectation is this only works for someone with a weak attachment to their personal style.
Your third option is to use, but also dispose of the rewrites. Instead of trying to have your style in the rewrite, use the rewrite as a learning tool you’ll dispose of after mining it for opportunities. Like how it was structured, rewrite your original draft in a similar structure. Like a particular phrasing? Copy that part back to your original draft. And overall, take in the feel of the rewrite. Does anything from the standard style of the AI tool appeal to you? As an example, maybe you do like some part of that direct confident style. Maybe you don’t want that entirely, but want some part of it.
Step 3: The Copy Editor, The Final Polish
It's usually best to get the structure and content right before getting picky about the details. However, you can use the AI as a copy editor throughout the process to clean up basic errors.
A word of caution: copy editing is not one of AI's greatest strengths. A simple grammar checker often does better, and a human eye will always catch things an AI misses. Don’t skimp on this final step. Your goal is to go from one author to many readers, and you'll make a bad impression and distract them if you leave simple mistakes in the text.
Common Pitfalls and Alternative Approaches
My collaborative approach has evolved in response to the common, but often less effective, ways people use AI for writing.
The "Vibe Writing" Trap
I’ve got the impression that most people using a GenAI model for the first time try what some are calling “vibe writing.” They use a prompt like “Write me an 800-word article on {topic}” and then iterate from there. It’s not hard to see why this is attractive; it’s an almost zero-cost start. However, I think many negative impressions of GenAI writing are based on the generic-feeling output from this method. The advantage of avoiding vibe writing is the opportunity to better shape and expand an initial, unique idea.
Beyond the Idea Generator
Another common pattern is using AI as an idea generator for tasks like creating an outline or suggesting titles. Early models did passably well at these tasks, so much of what you read about using GenAI will focus on this pattern. These still work when you need them, but thinking of these as the limits of the technology will be very limiting.
About AI and Truth
When using Generative AI, you’ll often be reminded to double-check. That’s good advice, these tools are probabilistic, and when they generate content, they can invent things that sound reasonable, but are not true. There are many non-AI cases where you should double-check facts, assertions and statements, so the general concern isn’t unique to AI.
Sometimes critics expand the possibility of untruth to suggest AI doesn’t have the capacity to distinguish truth from fiction, and is dependent on humans for this. That statement has a nugget of truth, but it’s also factually wrong. This claim is often followed by the argument that humans have a unique advantage: direct access to the material world for feedback. The problem with this argument is that for many questions, humans lack direct access to the truth either. Instead, both humans and AI rely on a process of searching for consistency across vast amounts of information—whether from other people, texts, or observations.
Humans are still better at this, but we have our own failings in distinguishing truth from fiction. Ultimately, both humans and AI are engaged in the same struggle. Our senses don't solve the search for truth, but by reason and access to information. A trained AI has the information advantage and lacks human cognitive biases. Humans still retain superior raw computational power. We also don’t fully understand human brains, and so they likely have additional advantages we don’t yet understand.
Note: In the realm of training, you might argue computational clusters are equal or greater to human capacity. But this would be an unfair comparison. Training is intrinsic to human existence. Response generation doesn’t use a full training cluster, but fractions of one, or entirely separate smaller systems. It’s likely to be a while before we individually have access to computing power equal to our own minds. The remarkable power efficiency advantage of the human brain makes it unlikely we’ll see computing power equal to the earth’s 8 billion human minds any time soon. That said, the human mind also has many inefficiencies for particular types of reasoning. This has allowed us to surpass human computational capabilities in specific domains with significantly less computing power. The amount of computing power a calculator uses to perform 2+2=4 is perhaps billions of times less than what the human mind demands for the same purpose. The human mind must constantly multi-task, vision, sensory, emotions. Ultimately it’s a very hard comparison to make.
If AI tools couldn’t distinguish truth from fiction, there’d be no purpose in asking AI tools to check your writing for accuracy, but you’ll find them quite capable of that. AI tools are getting better at avoiding untruth, and it is often by a deeper application of a search for consistency that they discover their own errors.
Still, they can miss their own errors, and yours too, so it’s always good to review yourself. The type of errors an AI will miss, is often different than the type you would miss. So your review is not just additional, but unique. As a final motivation, if you’re to be forgiven any errors, it would be those hardest for you. If you skip your review, and allow an error that would be easy for you to find, but is an AI weak point, you’re not likely to be forgiven.
Conclusion
The most effective way to use AI in writing is to treat it as a collaborator, not a vending machine. By moving beyond simple prompts and engaging the model in a conversation, you can leverage it as a developmental editor, a co-author, and a copy editor. The key is the interactive process. By asking it to identify weaknesses, justify its changes, and explore ideas with you, you transform it from a tool that generates text into a partner that helps you think. This collaborative approach doesn't just improve your writing; it deepens your ideas.
Author's Note: The creation of this article was its own journey of AI collaboration. I’m tempted to post examples of the writing process for this article. Would that be interesting? Posting here would have made this very long, so I’ll have to find alternative methods. I could post the link to the Gemini session, or screenshots. I have two Google Docs, one for the first draft, and a second from a midpoint. It would be a bit of work, but if the interest is there I would follow-up. I could also work on a more complete “prompt toolkit” for convenience's sake. I tend to think you should stay creative with prompting, but some process and repeatability can be useful too.