Drafting copy that doesn't read like AI wrote it

The AI-tell catalogue

Before you can fix it, you have to be able to see it. There are roughly five tells that mark a piece of writing as generated rather than written, and once you can name them you will see them everywhere — in product descriptions, in LinkedIn posts, in the captions under the year-end appeal from the INGO your cousin works at.

The hedged opener. “In today’s fast-paced world,” “in the rapidly evolving landscape of,” “as the digital ecosystem continues to mature.” These openings say nothing. They exist because the model, asked to begin, reaches for the safest possible noise. A human writer who knows what they are talking about begins with a specific thing — a number, a name, a customer who did a specific thing on a specific day.

The everything-and-nothing claim. “Our solution helps businesses unlock their potential and drive meaningful outcomes.” Read it twice. What does the product do? You cannot tell. The sentence has been written so carefully around the specifics that the specifics have evaporated. This happens when the model is given a brand name and asked to write copy without being told what the brand actually does, for whom, and why.

The three-sentence rhythm. AI prose often falls into a near-musical regularity: one medium sentence, one slightly longer sentence, one closing sentence the same length as the first. Paragraph after paragraph of it. Human writing breaks rhythm — a one-word sentence, a long winding clause, a question. The model will not do this on its own. You have to demand it.

The soft hedge. “Can help.” “May support.” “Could enable.” The model hedges by default because hedging is the safest output. The marketer writing the same claim with conviction says: does, will, cuts the time from four hours to forty minutes. The hedge survives because nobody told the model to drop it.

The corporate-blue tone. Every paragraph reads as if it were written by a mid-tier management consultant in a navy suit. Polite. Professional. Devoid of personality. This is the model’s safe default and it is the wrong default for almost every Nepali brand worth writing for. Khalti is not a navy suit. Daraz is not a navy suit. The local cooperative running a women-led savings group is definitely not a navy suit.

The fact-first, voice-strict prompt pattern

The fix is not to write a longer prompt. It is to write a structured one. Five lines, every time, before you ask for a single sentence of copy.

Voice: three concrete adjectives the brand IS — and two adjectives the brand is NEVER. Audience: one sentence — who they are, what they already know, what they want. Goal: one specific action you want the reader to take. Length: a number of words. Not “short”. A number. Avoid: three exact filler phrases the model is forbidden from using.

That is the entire pattern. Paste it into Claude or ChatGPT, fill in the five blanks, then ask for the draft. The output changes more from this five-line discipline than from any prompt-engineering trick you have read about online.

A real version, for a Khalti referral promo:

Voice: warm, plain-spoken, slightly cheeky — NEVER corporate, NEVER hedged. Audience: a 26-year-old in Lalitpur who already uses Khalti to pay her electricity bill and has three friends who don’t. Goal: get her to send a referral link to one specific friend today. Length: 60 words. Avoid: “seamless”, “empower”, “in today’s fast-paced”.

The output, with that prompt, is unrecognisable from the output you would have got with “write a Khalti referral caption”. The voice has a posture. The audience has a face. The goal has a clock on it. The forbidden phrases are gone before they had a chance to arrive.

The brutal editing pass

Even with the five-line block, the first draft is not the draft you ship. The first draft is the draft you cut. The brutal editing pass is a habit, not a step — once you have it, you will apply it to every AI draft for the rest of your career.

The rule is simple. Every sentence that does not say something specific gets cut. Not rewritten. Cut. If the sentence could appear, unchanged, in the marketing copy of a Singaporean SaaS company, a Mumbai fintech, and a Kathmandu NGO, then it is saying nothing and it goes. What remains will be shorter, sharper, and recognisable as written by a particular brand for a particular reader.

The hedges become claims. “Khalti can help you save time on bill payments” becomes “Pay your NEA bill in 12 seconds — not the 40 minutes the queue takes.” The first is true of almost any payment app; the second is true of Khalti, on a specific bill, with a specific saving. The reader believes the second because it costs the writer something to make.

The rhythm gets varied deliberately. Read the draft out loud. If three sentences in a row are the same length, break one. Add a one-word sentence. Cut a clause. The reader’s ear notices before their brain does.

The brand voice gets policed. If your brand never uses the word “empower” and the draft uses it twice, both come out. If your brand is plain-spoken and the draft uses “leverage” as a verb, that comes out too. The five-line block makes this easier, but it does not eliminate the need for a final read. The model is a chameleon, but it occasionally forgets which colour you asked for.

Three concrete worked examples

The Khalti referral promo. Fact-first block: a user who already pays her electricity bill on Khalti, three friends who don’t, a Rs. 50 referral bonus, the goal is one share today. Voice: warm, cheeky, plain. Avoid: “seamless”, “empower”, any English jargon. The model produces a 60-word caption that opens with a specific friend (“तपाईंको त्यो साथी जसले अझै लाइनमा बस्छ”) and ends with a clear ask: send the link, get Rs. 50, both of you. No mention of an “ecosystem”. No “fast-paced”. No corporate blue.

The INGO year-end appeal. Fact-first block: 1,247 girls supported in school in Sindhupalchok this year, the donor-retention rate is 38 percent, the goal is to convert last year’s donors into recurring monthly donors. Voice: respectful, specific, never sentimental, never “saving lives” generic. Avoid: “impact”, “transform”, “make a difference”. The model produces an appeal that names a girl, a village, a specific cost (Rs. 1,800 a month covers one girl’s school year), and an ask that is not “donate” but “switch your one-time gift to monthly”. The result is an email that reads like someone wrote it because they know the work, not because December is ending.

The B2B SaaS product update. Fact-first block: a 14-percent cut in average ticket-resolution time for the customer-support module, the audience is operations managers at Nepali ISPs and banks, the goal is to book a fifteen-minute demo. Voice: direct, technical, never “exciting news”, never “thrilled to announce”. Avoid: “revolutionary”, “game-changing”, “exciting”. The model produces an update that opens with the number, names the two specific customers who saw it, and ends with a calendar link and one sentence about what the demo will and will not cover. Three short paragraphs, no thrill, all ask.

The 90-second rule

Every AI draft gets a minimum of ninety seconds of brutal editing before it ships. Not because ninety seconds is magic — because it is the minimum honest time required to read the draft out loud, cut the empty sentences, replace the hedges with claims, and check that the voice still matches the brand. If your draft cannot survive ninety seconds of that, it was not ready to ship. Send it back to the model with the things you cut and a sharper avoid list, and try again.

The marketers who use AI well are not the ones with the longest prompts. They are the ones who refuse to ship anything that has not survived the ninety seconds.

Check your understanding

Quick check

—

You read a piece of marketing copy that opens with “In today's fast-paced digital landscape, our solution empowers businesses to unlock their potential.” Which tell of AI-generated writing is doing the most work to flag this as machine-written?

1. AIt uses too many long words.
2. BIt pairs a hedged, generic opener with an everything-and-nothing claim — the sentence could describe a payment app, a SaaS tool, an NGO, or a cooperative without changing a word.
3. CIt contains the word “digital”, which is a banned word.
4. DIt is too short.

Check answer→

Quick check

—

You have one minute before a campaign goes out and you want the single highest-leverage prompt pattern to make the model produce copy that does not read like AI. Which block should you paste in?

1. AA long paragraph describing your company’s mission and values.
2. BVoice (3 concrete adjectives the brand IS + 2 the brand is NEVER); audience (one sentence); goal (one specific action); length (a number of words); avoid (three forbidden filler phrases). Then ask for the draft.
3. CJust paste your old caption and ask the model to “make it better”.
4. DIncrease the temperature setting and hope for more creative output.

Check answer→

What comes next

A 60-word caption is one thing. A 1,500-word blog post or a four-page whitepaper is another — and the prompt patterns that win for short copy do not, on their own, win for long-form. The next section is about how to actually use AI for substantial writing without ending up with a fluent, repetitive, abstract piece that nobody finishes.