Most writers blame the model when AI output reads flat. The model is half the problem. The other half is the prompt. A bare instruction like "write a 500-word essay on remote work" pulls the model into its default register, which is the flattest possible version of professional English, and every detector on the planet is trained on that register. Better prompts push the model off that default before it produces a single sentence. Inside: a five-step prompt-craft workflow that puts audience and voice into the prompt before any content request, five techniques (persona prompts, voice samples, structure constraints, example-based prompts, anti-patterns) that each move the score on a different axis, a before-and-after prompt example you can copy today, why prompt quality matters more than model choice, how to use TextSight to verify the output landed natural, and a short FAQ on the honest tradeoffs of upstream control. Same model, very different output, depending on the prompt.
The order matters. Define the audience before the topic so the register has a target. Specify the voice before the content so the rhythm has a model to copy. Drop in example sentences before the ask so the cadence is set. Request varied sentence rhythm so the burstiness is built in. Demand specifics over generics so the prose has texture. Skipping any of these is the most common reason a thoughtful writer still gets flat AI output.
Open the prompt with who the writing is for, not what the topic is. A first-year college student who has read one textbook chapter. A skeptical buyer comparing two SaaS tools on a Sunday evening. A friend you have not seen in six months who asked what you are working on. The audience constrains vocabulary, depth, and tone in one line, and the model has read enough writing aimed at each of those audiences to know what they expect. Without an audience, the model defaults to the median professional reader, which is exactly the register every detector is trained on. Pick the reader before you pick the topic, and you have already moved the output a long way off the average.
Tell the model who is speaking and in what register. A tired but kind high-school teacher explaining this over coffee. A senior engineer reviewing a junior's pull request: direct, specific, no padding. A grad student who has just read three contradictory papers and is unsure which is right. The voice sets sentence length, hedge words, and emotional pitch in one sentence, and the model has enough fiction and interview data to know what each persona sounds like. The most common mistake is to leave the voice implicit; the model fills the gap with its flat default, which is what produces the recognisable AI cadence in the first place.
Paste two or three sentences in the voice you want, then ask for the new content in the same register. Examples carry more signal than any adjective. A persona is one line; an example is dozens of choices about rhythm, vocabulary, hedge words, and sentence opener variety. The model copies what it sees, which is what makes the technique so effective. Use rough sentences from a personal email or a Slack post rather than polished published prose, because the rougher source produces a closer match to your natural voice. This single step does more than the other four combined for most writers.
Request one short sentence under eight words and one long sentence over twenty-five words in every paragraph. Uniform rhythm is the most common reason human-edited AI prose still flags as AI, because the model defaults to a tight 16-to-22 word range across topics. Asking for the variance upstream is cheaper than fixing it downstream with an AI rewriter. Verify the rhythm in the output and discard any draft where every sentence lands in the same length band. The contrast between short and long is what carries voice at the sentence level, and it is the single mechanical fix that moves the score most reliably.
Tell the model that every claim must be grounded in a concrete example, a number, or a named source. Generic "many writers struggle with editing" must become "the three freelancers in my Tuesday writing group all skip the second pass." Generic "the meeting was unproductive" must become "we spent 45 minutes debating font choices for a slide deck that would not exist in six months." Specifics break the AI signature at the structural level because the model cannot invent grounded details; it can only synthesise them from the prompt. A piece with five concrete details usually reads natural at every level, even before the rhythm work, and it reads better to a human too.
Free includes 3 detector scans a day and a 1,500-word AI rewriter quota for the residual polish step after a stacked prompt. Paid tiers raise the quotas and add the Chrome extension, file upload, and REST API. Yearly billing saves 25%.
Billed $89.88/year — Save $30
Billed $179.88/year — Save $60
Billed $359.88/year — Save $120
Yearly billing saves 25%. View full pricing
Each technique moves the output on a different axis. Persona shifts vocabulary and tone. Voice samples shift rhythm and hedge words. Structure constraints shift cadence. Example-based prompts shift everything at once. Anti-patterns strip the words the model would reach for by default. Use one for a quick win; stack all five when the draft has to read natural at every level.
Open with a persona instead of a topic. A persona is a one-sentence character with an emotion and a context, not a job title. "A tired but kind teacher at the end of a long week" gives the model something to do; "a teacher" does not. The persona constrains vocabulary, sentence length, and emotional register in one line, which a long instruction list cannot match. Keep five or six personas in a snippet file (the tired teacher, the uncertain grad student, the senior engineer, the skeptical journalist, the patient older sibling) and pick whichever fits the voice of the piece.
Paste two or three sentences in the voice you want, then ask for the new content in the same register. Voice samples carry rhythm, hedge words, and quirks that no adjective can describe. The trick is to use rough samples, not polished published prose, because the rougher source produces a closer match. A line from a personal email, a Slack post, or an unsent draft works better than a magazine paragraph. Models tend to "improve" what they see; add the explicit instruction "do not improve the style, match the rhythm and the hedge words" to prevent that drift.
Stack mechanical rules. Use contractions throughout. Keep paragraphs to six short sentences. Average sentence length under twelve words. One short sentence under eight words per paragraph. One long sentence over twenty-five words per paragraph. No bullet lists. No headings unless asked. Each rule on its own is small. Together they break the default cadence the model wants to produce, which is most of what makes AI prose recognisable. Verify the constraints in the output and discard any draft that breaks more than one of them.
The combination of a voice sample and a worked example does more than either alone. Paste two paragraphs in the voice. Paste one paragraph of the kind of output you want, on a different topic. Then ask for the new content in the same combined style. The model now has both a rhythm to copy and a structural template to follow, which is the closest a single prompt can get to a fine-tuned model. This is the most powerful technique on the list and the one most writers skip because it requires both a voice sample and a worked example on hand.
List the words and constructions the model must not use. "Do not use delve, navigate, unlock, leverage, landscape, robust, holistic, seamless, multifaceted, intricate, tapestry, realm. Do not open a sentence with In today. Do not write It is important to note. Do not use parallel triplets." A banned-words list does more than ten positive instructions because models default to the same flagged vocabulary across topics. Keep the list short enough to remember on read-through, around twelve to fifteen words and three or four constructions, so you can verify violations without effort.
The cheapest demonstration of prompt quality is to run two prompts on the same topic with the same model and compare the outputs side by side. The first uses the default instruction style. The second stacks audience, voice, sample, rhythm, and anti-pattern. The difference is what every detector trained on the model's flat default can see.
"Write a 500-word essay on remote work." The model produces an opening about the rapidly evolving landscape of remote work, a middle section about leveraging collaboration tools to unlock productivity, and a conclusion about the multifaceted future of distributed teams. Every sentence lands between sixteen and twenty-two words. No specific examples, no named sources, no personal reference. Pasted into TextSight, this kind of draft typically lands in the 60 to 75 band on a 0 to 100 scale, with clustered highlights across the middle paragraphs. The detector is not wrong; the prose is the model's default register, and that register is what the detector was built to find.
"Write for a remote-work skeptic who has been forced back to the office twice in the last year. Voice: a tired but honest middle manager talking to a friend over coffee. Sample paragraphs: [paste two paragraphs from a personal email]. Structure: contractions throughout, one short sentence under eight words per paragraph, one long sentence over twenty-five words per paragraph, no bullet lists, no headings, paragraphs of four to six sentences. Specifics: every claim grounded in a concrete example or a number; mention at least two named tools and one personal anecdote. Do not use: delve, navigate, unlock, leverage, landscape, robust, holistic, seamless, multifaceted, intricate. Do not open with In today. Now write 500 words on remote work." Same model. The output usually lands in the 35 to 50 band, with scattered residual highlights rather than clusters. Two of those residual sentences can be cleaned manually or with the AI rewriter in Light mode.
The model did not change. The prompt did. The first prompt asked for an essay and got the median essay; the second prompt asked for a person and got something closer to one. Most writers spend their energy comparing models when the bigger lever is upstream of the model. A well-stacked prompt on a cheap model usually beats a bare prompt on the most expensive model, because the cheap model's default register is the same as the expensive one's. The flatness is in the training data, not the parameter count.
Writers spend hours comparing models and minutes writing the prompt. The ratio is backwards. Most of the quality gap between AI drafts in 2026 is upstream of the model, not inside it, because the default register is shared across every model trained on the same internet.
Every general-purpose model in 2026 has been trained on a similar mix of articles, essays, marketing copy, and Reddit posts. The flat professional register that detectors are built to find is the average of that mix, and the average is what a bare prompt pulls. Switching models rarely changes the underlying register; it changes the surface polish. A bare prompt on a state-of-the-art model produces a slightly nicer version of the same default cadence, which is exactly what the detector is trained on. Upstream control via the prompt is the only durable way to move off that default.
A heavy edit pass can strip the most obvious tells from a flat draft, but it cannot add a point of view, a specific example, or the small rhythm choices that make a paragraph sound like one person wrote it. Those qualities have to be in the source. A better prompt produces a draft that already has some of them, which means the edit pass starts further along. The same hour of editing produces a much better result on a stacked-prompt draft than on a bare-prompt draft, because the editor is now polishing prose with structure rather than imposing structure on prose without any.
Prompt quality is not a hack to game detectors; it is a craft choice about whether the writing has texture. A well-prompted draft reads better to a human reader for the same reasons it scores better on a detector: it has concrete examples, varied rhythm, an emotional pitch, and an identifiable voice. The detector and the reader want the same thing. The prompt is where you give it to them.
The five techniques compound, but only if you know which ones worked on a given draft. TextSight is the signal in the prompt loop. Paste each output into the detector before you decide to keep it, and use the score as a delta between prompt versions rather than a verdict on the final draft.
Run the same content prompt twice. Once with the bare instruction. Once with the stacked prompt. Compare the Authenticity Scores. The gap tells you how much the stack was worth on this topic, in this voice, with this model. Repeat the test with one lever removed at a time (no persona, no anti-pattern, no example) and you build a personal map of which techniques matter most for the kind of writing you do. Three or four experiments in an hour are enough to settle most of the open questions about your own prompt style.
The overall score tells you whether the prompt landed; the sentence-level highlights tell you which lines leaked. If two sentences in a row are highlighted, the issue is usually rhythm uniformity in that paragraph, which means the structure constraint was too loose. If a single sentence at the start of a paragraph is highlighted, the issue is usually a default opener the anti-pattern list missed. If half a section is highlighted, the issue is usually that the example sample was too polished for the topic. The map is diagnostic, not just a grade.
3 detector scans a day and 1,500 AI rewriter words a day, with all three AI rewriter modes available, covers a heavy prompt-iteration session at no cost. Pro at $14.99 on annual ($19.99 monthly) removes the daily caps and adds 50,000 AI rewriter words a month, file and URL upload, and the Chrome extension for in-context scans while you write. The free tier is enough to learn the loop; the paid tier is for the writers who run the loop daily.
If you wrote a stacked prompt and a few sentences still flag on the scan, the TextSight AI rewriter can resolve them without rewriting the draft. Pick the lightest mode that works. Heavy rewrites on prompt-crafted prose usually make it worse, because they replace the voice you just shaped.
Light keeps the prose close to the original and is the right starting mode for output from a stacked prompt. Use it on the two or three sentences that flagged on the scan, not on the whole draft. Light typically moves a flagged sentence score by fifteen to twenty-five points without changing the meaning. If the rewrite drifts even slightly from what you meant, reject it and edit the sentence by hand instead. The AI rewriter is a polishing tool at this stage, not a rewrite engine.
Standard rewrites more aggressively and is appropriate when a paragraph from the prompt output still flags after a manual pass, usually because of uniform sentence rhythm the structure constraint did not fully break. Standard handles the rhythm fix more reliably than Light at the cost of a slightly less faithful echo of the prompted voice. Use it on one paragraph at a time, not on the full draft. After the rewrite, read the paragraph aloud and accept only what reads as the voice you set in the prompt.
Maximum is built for heavy AI drafts where the prose was the model's default register from the start. Running Maximum on prompt-crafted output usually makes it worse, because it replaces the voice you shaped with the AI rewriter's idea of natural prose. The two times Maximum makes sense on this work are translation polish (where the prompt was in a non-English source) and emergency salvage on a deadline. In both cases, a manual read-through afterwards is non-negotiable.
The assistant-mode workflow that keeps prose yours while AI does outlines, research, and grammar polish.
Read the assistant-mode guideFive rewrite tactics for the days the prompt output still needs a craft pass on the prose layer.
Read the rewrite guideHow the 0-to-100 metric is computed and what each tier means for prompt iteration.
Read the guideThe standalone AI rewriter tool for the residual sentences that survive a stacked prompt.
Open the AI rewriterDetector, AI rewriter, and sentence-level highlights in one workflow. Free to try with no card. 3 detector scans and 1,500 AI rewriter words on the free tier, every day, so the prompt loop is free to learn.