AI makes writing for fun even more fun and writing for work bearable.
My philosophy on using AI for writing is simple: it will either make my writing a lot better without saving time, or it will save time and make my writing a lot worse. The choice is one or the other. I pick better.
As of mid-2025, there’s still a stigma about writing with AI. But the future here is obvious. The old way of writing, where humans craft and edit every sentence from scratch, is ending. AI will do a lot of the heavy lifting, converting human ideas into text. Young people are already embracing AI as an essential tool. There’s zero chance they use AI for everything in life except writing.
A quick note before I get into the process: always use the best models available. They follow instructions more accurately and produce better output at every stage. The difference between a frontier model and last year’s version is enormous, especially when it comes to following a style guide or organizing a complex argument.
Here’s my approach.
Build a style guide first
Before writing anything, I have a personal style guide ready. It’s a set of custom instructions that I feed to the AI at every stage of the process.
Here’s how I built mine. I found authors whose writing I admire, fed a few of their pieces into the best model I had access to, and asked it to analyze the style. Then the AI generated a set of writing instructions that replicate that voice. Reading level was important to get right. I aim for 6th to 8th grade because that’s where writing feels effortless to read, and effortless reading means people actually finish the piece.
My style guide goes beyond grammar. It includes rules about sentence length, word choice, tone, and a long list of patterns to avoid. Preferences about argument structure and hedging are in there too. The more specific the guide, the less editing I need to do later.
In Claude, this lives as a skill file, then a project with custom instructions to use that skill when writing. In ChatGPT, the custom instructions feature in a project or a dedicated Canvas works well. Either way, the style guide is a living document. Every time I notice a new pattern I want to add or remove, it gets updated.
Determine the approach
Every article starts with a decision about form. Am I writing an essay? A listicle? A how-to guide? An opinion column? The form shapes everything that follows, so I decide early.
Then I describe the article to the AI: the argument, the intended takeaways, the approach. This description becomes the foundation for the whole process, so thoroughness matters here. More context upfront means less time correcting course later.
Write a stream of consciousness
Everything starts with dumping ideas onto the page. Grammar, order, completeness, none of it matters yet. The only goal is getting everything out of my head and into a document. If I have the luxury of time, I leave it and come back when something new occurs to me. Once I’ve said everything I have to say, I move on.
This step matters because it ensures the core ideas are mine. AI is good at organizing and polishing, but the raw material should come from my own thinking whenever possible.
Ideation
The first AI step is organizing my stream of consciousness into a rough outline. I tell it to ensure the outline flows coherently from one topic to the next. If it works properly, AI gives me a structured outline using my ideas and only my ideas.
Then the outline becomes a jumping-off point. I tell AI the goal of my article and ask what other topics I should cover that aren’t already included. It spits out a list of ideas. Some are obvious, some are genuinely useful, and some are filler. The ones that resonate get a follow-up. I ask it to go deeper on a specific suggestion, or to give me a few more ideas related to a particular subtopic.
This loop runs as many times as it needs to. If a suggestion requires research, I feed the AI relevant documents or articles to work from. Then it’s back to the stream of consciousness to write more based on whatever new ideas fit with the piece. The goal is to end this stage with more raw material than I started with, all of it filtered through my own judgment about what belongs.
Build the outline
With expanded notes in hand, I regenerate the outline and start pushing on it.
This is where the process gets more deliberate. I ask the AI to mark where in the outline it will use examples, cite data, or reference sources. This forces the outline to be concrete rather than vague. Then it provides sources for those examples and data points, and I check every single one. AI hallucinates sources regularly. Fake citations get removed, or I find a real source myself.
Once the outline feels solid, the AI reviews the order of information. I challenge it to organize everything so the reader has all the context they need as they encounter each new idea. The structure should be persuasive, because the order in which information is presented is half the argument.
One thing to watch for: how much space the outline gives each topic. AI likes to cover every section equally, and I usually want some topics expanded and others condensed. AI gets me 80 to 90 percent of the way to a complete outline. After that, I copy it into a separate document and push it over the finish line by hand.
First draft
The style guide and completed outline get combined into a single prompt. Something like “Write a 1,200-word article based on this outline” works well.
Then comes heavy editing. I read through the draft and selectively ask AI to revise specific paragraphs. Thin sections get expanded, but only with information I provide so AI doesn’t fill gaps with fluff. Belabored points get cut.
Good writing uses examples, and this is a major weakness of AI. The examples tend to be overused, hallucinated, or missing entirely. The same goes for quotes. My own go in where necessary.
Repetition is another problem to watch for. AI won’t reuse the same sentence, but it will make the same point in slightly different language across multiple paragraphs. Since I’ve written other articles with AI’s help, crossover between pieces is a concern too. AI has a set of favorite talking points for any given topic and will recycle them unless actively pushed away.
Second draft
At this point my edits have probably made a mess of the first draft. The whole thing goes back through AI, warts and all, with a simple prompt like “Clean this up.” This gives AI a chance to smooth out the seams between my manual edits and its original prose, and the result is a much more cohesive piece to work from.
Sometimes I experiment with asking it to rewrite the article in a specific style, like “Rewrite this in the style of a New York Times opinion article.” Either way, I only proceed when the article is, in my view, 90 percent done.
Fact-check everything
The AI fact-checks itself by searching the web. All sources get cited, and then I check every one of them. This step is tedious and essential. AI will confidently present fabricated statistics and misattributed quotes. Skipping this step means publishing someone else’s mistakes under my name.
Edit out the AI tells
This is the stage where I earn my byline. AI has recognizable writing patterns, and leaving them in signals to any careful reader that the piece was machine-generated. Some tells can be fixed globally with a single prompt. Others require manual editing.
The easy fixes come first. Removing all em dashes, for instance. ChatGPT loves em dashes, but real people rarely use them. Heavier copy editing works well here too: changing all instances of passive voice to active voice, or rewriting the article in AP style. These kinds of global passes are where AI editing is genuinely faster than doing it by hand.
Then the article gets exported to a document I can edit by hand. Here’s what I look for.
Contrast framing. AI loves constructions like “This isn’t just about growth, it’s about trust” or “It’s more than a tool, it’s a partner.” These feel punchy in isolation, but AI uses them so often they’ve become a calling card.
Fake depth language. Phrases like “at its core,” “fundamentally,” “in today’s fast-paced world,” “the reality is,” and “it’s worth mentioning” sound meaningful without actually saying anything. They get cut.
Too much symmetry. AI produces sentences that match each other in length and rhythm constantly. Humans do this occasionally for effect. Two consecutive sentences that feel engineered to complement each other means one of them needs rewriting.
Universal neatness. Every paragraph about the same size, every thought landing cleanly, no weird detours or awkward emphasis. Real writing is a little messy. An article that reads like a completed outline needs more human touch.
Overexplaining transitions. “That said,” “with that in mind,” “in other words,” “meanwhile,” “when it comes to.” AI glues every paragraph together too carefully. Sometimes the best transition is just starting the next paragraph.
Generic motivational tone. Even in business writing, AI sneaks in uplift and reassurance. “By taking a thoughtful approach, teams can unlock new opportunities.” Sentences like that get deleted or rewritten with a concrete claim.
Excessive hedging. “Can be,” “may help,” “often,” “typically,” “in many cases.” Some hedging is normal and honest. AI stacks it until the writing loses all force.
No real stake in the argument. AI can sound confident, but it often avoids committing to a concrete judgment. It summarizes when it should be taking a stand. An article that never actually says what it thinks needs more of my voice in it.
Thesaurus drift. Words that are technically correct but slightly off for the context. “Leverage” where “use” would sound normal. “Robust” where a human would just say “strong.” Every word that isn’t in my normal vocabulary gets swapped.
Repetitive sentence openings. “This means…” “This can help…” “By doing so…” “Whether you’re…” Once I noticed this pattern, it started showing up everywhere.
Too much signposting. “Here are several key reasons why…” followed by paragraphs that deliver exactly as promised, each one about the same length. Real people often drift a little, spend more time on the interesting parts, and skip the boring ones.
And the one tell I can never fully fix: AI writing betrays no lived experience. There are no real stories, no genuine surprise at what the writer found, no moments where someone clearly got interested in a tangent. Those have to come from me. That’s what makes the article mine.
One trick for the manual editing stage: when a paragraph reads like AI wrote it but I can’t quite fix it myself, I copy and paste it back into the AI with specific instructions about what to change. Sometimes the fastest way to make AI sound human is to make it rewrite its own worst habits.
Final draft
The last pass is by hand. Final touches, personal quirks and preferences, the small things that give writing personality. Then the entire text goes through an AI proofreader for typos and grammar.
After that, one question. If I publish this and it’s criticized, who gets the blame? The answer is me, obviously. But if I’d have even the slightest temptation to deflect any part of that criticism onto AI, I’m not done. The work continues until I feel like I own every sentence, even the imperfect ones.