Ask a language model for an article and you will usually get something competent. The structure is clean, the paragraphs are balanced, the grammar is flawless, and somewhere around the third paragraph you notice the same rhythm repeating: medium sentence, medium sentence, medium sentence. The content reads like a summary of an article rather than an article someone wrote. If you publish that on its own, readers will notice, sometimes without being able to say why. The fix is not a longer prompt, and it is not a request to “write more humanly.” It is a set of specific, mechanical instructions about voice, rhythm, and structure that you can reuse.
Start by Defining Who Is Speaking
The single biggest reason AI writing sounds generic is that nobody told the model who is speaking. A default request produces the model’s own average voice, which is a compromise across everything it was trained on and lands nowhere. Anthropic’s guidance for prompt engineering starts from the same place: always begin by clearly describing the task, and treat the model like an intern on their first day, giving clear, explicit instructions with all the necessary detail. The same applies to voice.
Skip vague instructions like “write in a conversational tone” or “make it engaging.” They produce vague results because they describe a category rather than a person. Replace them with a description specific enough that you could tell two different writers apart. “Write like an experienced trade blogger who explains practical steps without oversimplifying or sounding academic” tells the model something it can act on. “Write like a veterinarian with fifteen years of clinic experience who has seen owners ignore the same advice three times” tells it something tighter still. The more concrete your description of the speaker, the less room the model has to fall back on its default register.
Give It a Physical Point of View
Voice gets sharper when it is attached to a specific vantage point. Instead of “write about sleep quality,” try “write about sleep quality from the perspective of someone who has spent a decade debugging their own insomnia and is tired of generic advice.” Now every paragraph has a filter to pass through, and the model has to make choices a general article would not require. That constraint is what removes the smoothness.
Use Examples Instead of Adjectives
Few-shot prompting is the technique Anthropic highlights for teaching a model a specific format or style: you provide realistic examples of the inputs and the ideal outputs you want. Their published examples include deliberately difficult cases and edge cases, because those are the ones that actually pin down the boundary of what you mean.
This is the most reliable fix for flat AI prose, and it is faster than it sounds. Two or three short samples of your own writing, pasted straight into the prompt, will steer tone more effectively than any adjective you can invent. The samples should show the specific things you want more of. If your problem is stiff openings, include an opening that starts mid-thought. If your problem is list-heavy structure, include a paragraph that weaves two points together without a bullet in sight. If you want the model to use contractions and drop the occasional sentence fragment, show that in the sample rather than requesting it.
Break the Task Into Stages
Anthropic also recommends prompt chaining for complex work, breaking a task into a sequence of prompts where each new prompt can include the previous prompt-response pairs. The same logic applies to writing an article. Asking a model to produce a finished 1,500-word piece in one shot is asking it to make dozens of decisions simultaneously, and the ones that shape voice are the first to get averaged away.
Ask for an outline first. Then ask for the opening section, and feed the result back in for the next one. Finish with a separate editing pass that reviews the assembled draft against your voice brief. Splitting the work slows you down slightly and removes the uniformity that makes AI text obvious, because each stage has one job.
Break the Rhythm on Purpose
The most commonly cited tells of generated writing are structural rather than lexical. Sentence length does not vary. Paragraphs are the same size. Transitions do the same job in the same place every time: “Furthermore,” “Moreover,” “Additionally,” “In conclusion,” “It is important to note.” This vocabulary blacklist is a symptom, not the disease, and removing the words without changing the pattern leaves the problem intact.
Give the model an instruction about rhythm instead. Ask it to mix short declarative sentences with longer explanatory ones, to start at least one paragraph without a linking word, and to let some paragraphs run long while others stay to two lines. Ask it to cut filler that models add by default. Then read the draft aloud. The test that matters is not a detector score, it is whether your ear catches a rhythm that stays the same for a page.
Replace Generic Claims With Specific Ones
Generated drafts tend to make claims that sound reasonable and could apply to any topic. “This approach can significantly improve your workflow” belongs to every article ever written. Search your draft for sentences that would still be true if you changed the subject, and replace them with something concrete: a number you can source, a limitation, a step that costs time, a case where the advice does not apply.
This is where the biggest quality jump happens, and it does not depend on the prompt at all. A model cannot invent your proprietary data or your actual experience. It can, however, hold a frame open for them if you ask. Tell it to leave a specific, clearly marked slot for a real example, and treat that slot as a reminder that the piece needs evidence you have to supply.
Add a Constraints Block at the End
Both OpenAI and Anthropic recommend placing critical instructions where the model will actually attend to them, and long-context research from Anthropic specifically recommends putting key instructions at the end of a long prompt. Practically, this means your constraints belong after the material, not buried in the opening paragraph.
A closing block of four or five rules is enough. “No transition phrases such as furthermore or moreover. No sentence longer than 30 words. No paragraph longer than four sentences. No conclusion paragraph that summarizes what was just said. Write like a practitioner, not a summarizer.” Read it back before you send, because a constraint you cannot state clearly is a constraint the model cannot follow.
A Reusable Base Prompt
Here is a starting structure you can keep in a text file and modify per project. It is deliberately plain, because the value is in the specificity you add rather than in the format.
Write a practical guide for independent content creators. The speaker is an experienced practitioner who has done this for years and is tired of generic advice. Voice: direct, specific, willing to disagree with the obvious answer. Structure: no summary introduction, no conclusion that recaps the section above. Rhythm: mix short and long sentences, vary paragraph length, never begin two consecutive paragraphs the same way. Banned: furthermore, moreover, additionally, delve, landscape, in today’s fast-paced world, it is important to note. Every claim must be concrete and checkable; if a specific example is needed, insert a clearly marked placeholder rather than inventing one. Ask questions when you need information instead of guessing.
Then Edit Like an Editor, Not a Proofreader
Prompting gets you 70 percent of the way. The last 30 percent is a human pass, and it should be an editing pass rather than a proofreading pass. Read for rhythm first, cutting or splitting sentences until the pace changes from paragraph to paragraph. Then read for specificity, replacing every claim that could be swapped for a different topic. Only then fix the grammar, because polished grammar on a flat argument just makes the flatness harder to spot.
Two habits make the biggest difference. Keep a running list of the phrases you personally dislike, and add them to your banned list so every future draft starts without them. And save your best prompts as reusable blocks, because the effort of writing a good voice brief is worth doing once rather than every time you need an article.
The goal is not to fool a detector, and chasing that goal leads to writing that is worse for readers. The goal is to produce text that carries a specific person’s judgment, which is exactly what the model lacks on its own. Give it a voice, give it examples, give it rhythm rules, and leave a hole for your real experience. The result reads like a person wrote it, because a person did.
