A Brief(ish) Actionable Guide on How to Write with AI Support
A stage-by-stage framework for writing better with AI.
This will be an unusual issue. While the data show that practically the only way to go viral or make money with a periodical like this is to write about self-improvement and listicles, I’ve stayed well clear of that content. But not today!
My first experience using AI at The Long Horizon was an abject failure. I thought I could upload my published content into memory and get AI to create new issues from a blank page for me to edit. I spent a few days trying to make this work, but the result was garbage.
That being said, the web tool I use for writing (Lex.page) invites experimentation with AI because of its streamlined integration. After tens of thousands of published words and countless experimentations with AI over that time, I’m still learning but have some rhythms to share.
However, this is still The Long Horizon. So, below you’ll find mostly framework-based content, serving as a foundation that you can adapt to meet your specific needs. This methodology isn’t limited to public writing; I use it professionally as well.
Part 1: A Framework
Using AI as a writing assistant comes down to two things:
Knowing when to use AI.
Knowing how to use AI.
And to make matters slightly more complex, the second proposition changes vis-à-vis the first. We end up with an ‘equation’ like this:
Success at Using AI for Writing = (Using at the Right Time) x (Using the Right Way for That Time)
To demystify that for practical use, here’s a rough approximation of how I use AI as a writing assistant (as a graph, because, why not?):
Here are my stipulative definitions for the compositional stages:
Ideation: Answers the main question “What should I write about?” and sub-questions around points or ideas that might be germane to the piece and (hopefully) interesting.
Raw Content Creation: Generation of sentences and paragraphs containing main claims, supporting claims, and random germane thoughts that may or may not survive into the final piece. A soup of content that engages with the what, why, and why it matters. Usually has some structure, but organizing is not the main focus.
Narrative, Coherence and Argument: The ‘meta’ of the piece. This stage focuses on how the piece reads, whether arguments are supported or unsupported, whether the through-line is clear or zigzags too much, whether reasonable counter-arguments are understood and addressed (or willingly left unaddressed), etc. The discovery of issues in this stage often necessitate going back to the Raw Content Creation stage.
Copy Editing: Focuses on aspects such as transitions between sections, repetitiveness or redundancy, clarity, active voice, author style/voice, readability, etc.
Finalization: Final checks on grammar, title, subtitle, links, etc.
The next section will provide a deeper dive into how I use AI support in each compositional stage.
Part 2: Deep Dive—Using AI Across Composition Stages
The key to effectively using AI for writing support is knowing the unique [How x When] combinations. The following subsections will give my impressions on how AI can support you during each stage, its strengths, weaknesses, and specific prompts I use.
This advice assumes you’ve loaded some contextual information to the platform1, including some/all of:
Target audience profile.
Your past work in the same domain, to indicate voice, style, and past information your audience may have consumed.
Goal of the piece (e.g., for an Op-Ed or professional memo), or of your writing in general (e.g., for a blog or column).
Supporting documentation, such as relevant research papers, articles, and adjacent knowledge.
I recommend starting there and then experimenting with the below.
A. Ideation
In the graphical representation, you’ll notice (a) I don’t recommend starting with AI, and (b) the AI intervention here is brief. I’m curious about others’ experiences, but when I ask AI for ‘bounded blue-sky’ ideation2 it either generates ideas too similar to my existing content, or that are too ‘out there’ to be useful.
However, AI excels at providing scaffolding to generate your content from and at unearthing ‘gaps’ to fill. If you provide a main idea and a few supporting ideas, it can generate an outline and suggest other points to consider working into the piece.
At this point in time, it is unwise to ask the AI to generate entire paragraphs or complete sentences. This is because (i) it can cause path dependency and prevent you from putting your true (even if rough) thoughts to paper (which, presumably, is why people will want to read what you write); (ii) it is unlikely to sound like you; and (iii) it often suffers from hallucination problems.
Example Prompts
“Here is my premise: ______. Do you think that will be interesting to my readers? Why or why not?”
“If my main hypothesis is ______, what kind of sub-claims would you expect to see?”
“Here is a main idea, and five other related points. Can you arrange them into a coherent outline? If you think there are missing points that are critical to the argument or ‘connecting the dots,’ please include those and highlight them for me.”
B. Raw Content Creation
Writers’ brains work differently. My approach is to dump thoughts onto paper (often discontinuously) and then refine, fill in gaps, and organize.
The downside of this style is that the end product of the Raw Content Creation stage lacks overall coherence and a narrative through-line, even while full of content. More often than not, there are entire sections that will, at worst, end up being deleted or, at best, be moved to an entirely different point in the piece. Thankfully, the next stage takes care of that. Before AI, that was quite laborious—but that’s where AI writing support shines (more on that in the next section).
Similar to the Ideation stage, I find that the AI intervention in Raw Content Creation is best kept limited and brief. At this point (i.e., February 2025), even when you’ve uploaded context, AI isn’t very good at generating raw content that has the trifecta of (i) follows the argument (ii) is sensible and (iii) sounds like the author.
However, it is good at powering through writer’s block! At this stage, the most useful tool in the AI arsenal is not asking it questions (as we are thoroughly trained to do with past technology—i.e., web searches), but instead ask it to ask you questions.
Lastly, in the same vein, AI support in this stage can help the ‘micro-blocks’ that are notorious for slowing down writing momentum—i.e., grasping for a word, phrase, concept, or source material. Before AI, these kinds of frustrations might spell the end of a writing episode and a task switch. Nowadays, I just ask AI, plug the hole, and move on (knowing that even if it is a 70% answer, it’ll get caught in later stages and then refined).
Example Prompts
“Please read through the section titled ______. If you were in discussion with me on this topic, what kind of questions does the current draft raise for you?”
“I’m stuck. Particularly, I know that I have not made a compelling argument, but it is unclear to me what I can reasonably fit in this piece that readers will find interesting, important, and germane. Please read through what we have, and then ask me three questions that ‘zoom out’ and might help me orient myself and get unstuck.”
“What underlying concepts might be needed to understand or support this argument: [Cut and paste claim]?”
“I’m looking for a word (or phrase) that means a little bit of _____, some of ______, and also ______.”
“Please finish this sentence ‘[Cut and paste fragment] …’”
“What would you label this section—i.e., the one that currently starts with ‘______’ and ends with ‘_____’? Next, please provide a brief explainer as to how you see that section fitting in the overall piece.”
“What stands out as missing from the section titled ‘_____’?”
C. Narrative, Coherence and Argument
If you take anything away from this guide, it should be that, as of this writing, this is where AI support really shines—what I call the ‘meta’ of a piece.
For me, content creation is the ‘fun part.’ Then, during this stage, fatigue sets in. If not dealt with properly, that fatigue, when combined with something psychologists call the Curse of Knowledge, can result in a piece that fails to deliver on the promise of the hypothesis and the raw ideas.
AI proves a reliable and effective sparring partner to combat this. A repeating process of ‘zooming out’ and then ‘zooming in’ works really well. For example, asking about the through-line of an argument, identifying the weakest link, getting help improving it from an ‘outsider’s’ point of view, and then repeating until satisfied. I start this stage by asking AI about strengths, weaknesses, and narrative flow. The first go-around of this usually provides a plethora of meta advice. If you are overwhelmed, you can rein in the advice by asking it for just one area that you need to work on (and then repeatedly do so).
Furthermore, this stage is when an entire section is most likely to be deleted (if it is no longer relevant) or moved. The same is true for a sentence that feels off.
If I am going to ask the AI to rate the piece, this is when I do that. I usually ask it to do so based on a few pre-defined dimensions (e.g., originality, strength of arguments, compellingness, interestingness).
As you see in the graphical representation, I spend as much, if not more, time in this stage than in raw content creation. In reality, I am moving back and forth between the two quite a bit.
Example Prompts
“Is anything I’ve written so far inaccurate, or even just misleading?”
“Am I conveying the ideas in this piece clearly?”
“Please read through this entire draft. Then, create an outline of the piece in the following format: (A) Section Title, followed by (i) Subpoint 1 - theme or idea, (ii) Subpoint 2 - theme or idea, etc. Next, please review that outline and suggest whether anything seems out of place or missing.”
“Please note this [cut and paste sentence(s) or paragraph(s)]. If you were going to move it to a more optimal location in the piece—i.e., where it better fits the narrative—where would that be?”
“Please read through this entire draft. What other source material should I, or my readers, consider when evaluating this claim or argument?”
“Please review the entirety of this piece and evaluate it, on a scale from 1 - 10. Specifically, I’d like you to consider dimensions such as _____, _____, _____.”
D. Copy Editing
Any writer who has ever worked with an editor, or even just a really adept, helpful colleague, has heard the following advice: “Kill your darlings.”
I extend this advice quite widely to include:
Things written early on in the creation phase that no longer serve their purpose.
Sentences that have literary heft but are out of place for the content or medium.
The writer trying to show how smart they are, or how much they know, regardless of whether it strengthens the piece.
And anything the writer ‘wrote for themself’ and not for the reader.
While these might seem obvious, they are everything but to a writer grappling with a piece. Thus, ‘editor’ exists as a profession. But not all of us are fancy enough to have a proper editor. That was a problem pre-AI. Post-AI, as long as you are abiding by the [How x When] framework, you can get stellar results from using an AI as your copy editor.
This stage is for tackling sentence structure, readability, redundancy, succinctness, phrasing, voice, grammar, and other optimizations. Notably, my preferred writing tool (Lex.page) has a feature that allows you to run pre-programmed checks to accomplish the goals of a copy editor, and I find they work quite well.
As you can tell from the graph, AI is actually doing more of the work in this stage than I am—mainly in the form of deletions!
Example Prompts
“Are there any parts of this draft that are really just fluff? Where it seems like I am bloviating? Where it seems like I am too attached to a phrase that really ought to be killed?”
“Please read through the documents in your memory. With those documents in mind, now read through the current draft. Does any language in this draft not sound like me?”
“Please read the entire current draft. Do you find any parts particularly redundant? If you do, please think about whether the redundancy is used to make an important point more powerfully, or if it could be removed.”
“In this paragraph, I want to keep this exact sentence as is: ________. That being said, the paragraph is suboptimal because it _____. Could you suggest a rewrite that preserves the sentence and the main idea?”
“In this paragraph, my goal is to ______. Do you think the paragraph succeeds? What ‘big picture’ suggestions do you have to improve it?”
“Will you please evaluate the transition between the section titled ______ and the one titled ______? Do you have any recommendations to smooth the transition?”
E. Finalization
In terms of finalization, I use AI to get over any reservations I have and gain the courage to just hit ‘send.’ No piece is ever perfect, but there is such a thing as overoptimizing.
Oftentimes, this is a good time to take a fresh look at things like your title, subtitles, and any footnotes. AI can help identify where footnotes and/or sources may be needed to further bolster your arguments. A final check on grammar, spelling, etc. also never hurts.
Example Prompts
“Please suggest 3 titles that you think hint at the nature of this piece, but that are also likely to seem compelling to my audience.”
“Please read through the entire draft one more time. Are there any spots where you strongly suggest I should consider including a footnote or a source?”
“Please suggest a short blurb that I can use to post this on social media. It should be [x] words or less.”
If the platform you are using doesn’t allow this (or it isn’t yet enabled because you aren’t paying for it), you are going to be handicapped from the beginning and disappointed with the results.
“Bounded” meaning that I instruct it to give me ideas that fit within a specific container—e.g., such as within the topical areas of The Long Horizon.