Letting My Machine Write First:
The blank page has always been the expensive part. Not the research, which is mostly patience and not the revision, which is mostly judgment, but that first descent from a vague intention into actual sentences on a screen. Anyone who writes regularly knows the particular procrastination it produces and anyone who writes for a living knows how much of a deadline it can consume. Large language models have changed the economics of that moment and it is worth describing plainly what they change and what they emphatically do not.
Start with the honest framing. What these systems produce on a first request is competent, fluent, structurally sound and almost never good enough to publish. It reads like the average of everything written on the topic, because in a meaningful sense that is what it is. Treating the output as finished work is the mistake that has flooded the internet with prose nobody wants to read and it is also a mistake that gives the whole practice a bad name among serious writers. Treating it instead as raw material, the equivalent of a quarry rather than a cathedral, puts the tool in its proper place and makes the resulting workflow genuinely powerful.
The most productive way to begin is to give the assistant far more context than feels necessary. A request to write about project management yields platitudes, a request that specifies the audience as engineering leads at companies under fifty people, the angle as skepticism about estimation rituals, the desired length, the voice and the three claims you already intend to make will yield something you can actually work with. The difference is not subtle and most disappointment with these systems traces directly to under-specified prompts. Think of it as briefing a capable freelancer who has read enormously but has never met you, knows nothing about your readers and will otherwise default to the safest possible interpretation of your request.
From there, the useful pattern is iterative rather than one-shot. Ask for an outline before prose, argue with the outline, then commission sections individually rather than requesting the whole piece at once, because quality degrades noticeably across long single generations and because reviewing in smaller units keeps you engaged rather than passive. For a book, this scales naturally: chapter summaries first, then a rough version of one chapter, then a critical pass where you ask the model to identify the weakest argument in what it just produced. That last request is oddly effective, since these systems are considerably better at criticism than at self-directed excellence.
Now the crucial part, which no amount of clever prompting replaces. Everything generated must be verified and the verification burden falls entirely on you. Factual claims may be confidently wrong, citations may refer to papers that do not exist, statistics may be plausible inventions and quotations attributed to real people may never have been said. The fluency of the output is precisely what makes these errors dangerous, because fluent text triggers the reader’s trust reflex, including in the reader who happens to be its editor. My own practice and the only defensible one I know, is to treat every specific assertion as unverified until checked against a primary source and to delete anything I cannot confirm rather than soften it into vagueness.
Voice presents a subtler problem than accuracy. Machine-generated prose has recognizable tics: the fondness for tricolons, the reflexive both-sidesing of any contested question, the transitions that announce themselves, the closing paragraph that summarizes what was just said instead of landing somewhere. Readers sense this even when they cannot name it and the sensation is one of reading something that nobody in particular wrote. Fixing it requires actual rewriting rather than light editing, which is why the honest accounting of time savings is more modest than the enthusiasts claim. What disappears is the friction of starting and the labor of structuring; what remains is the work of making the thing sound like a person with opinions.
Where does this leave the writer’s own thinking? This is the question worth taking seriously, because writing is not merely the transcription of finished thoughts but the process by which thoughts get finished. A first pass composed by someone else, machine or human, deprives you of that discovery and it is entirely possible to end up with a polished piece that expresses a position you never actually reasoned your way into. The defense I have found is to write the argument’s spine myself, in rough notes or a few unpolished paragraphs, before asking for any assistance at all. The model then elaborates and arranges material I have already thought through, rather than supplying conclusions I have not.
Certain kinds of content suit this collaboration better than others. Explanatory material, documentation, survey pieces and structured reference work benefit enormously, since their value lies in organization and completeness rather than in originality of perspective. Reported journalism, personal essay, memoir and anything whose worth depends on a specific person having seen or felt something do not benefit at all and attempting it produces the characteristic hollowness that readers reject. Technical writing sits somewhere in between, improved by the tool where it covers known ground and unaided where it describes what you personally built and discovered.
Disclosure deserves a paragraph of its own. Practices vary, professional norms are still forming and the right answer depends on context in ways that make blanket rules unhelpful. Academic and journalistic settings increasingly have explicit policies and violating them is a serious matter. In commercial content the expectation is looser, though clients reasonably want to know what they are paying for. My own inclination is toward transparency where the question would matter to the reader and toward the general principle that if you would be uncomfortable disclosing how a piece was made, that discomfort is information worth heeding.
My perspective, after considerable use, is that these systems are an excellent second chair and a poor first violinist. They eliminate the dread of the empty document, propose structures you can argue with, produce serviceable connective tissue and never tire of revision requests at eleven at night. They do not know what is true, they do not know what you think and they cannot tell whether a sentence lands. Held to that division of labor, the arrangement produces more writing and better writing than working alone; held to any grander claim, it produces exactly the sort of text everyone has already learned to skim past.
AI writes better than I do. I feel more like an editor than a writer. I’m more interested in seeking the truth and what things mean, than who gets credit for discovering a particular truth. I do like acknowledging sources. I like giving people credit. I like getting credit for my original ideas. I’m usually trying to learn about the subjects I am writing about. Editing AI responses is a great way to learn about things. And I figure the results will be valuable to readers seeking information about the subject I’m writing about. So, I ask AI a question about something I’m interested in and then, edit and polish the result into a story that I can understand. Then, I concatenate some of the stories into an article or book.