AI Content That Doesn't Sound Like AI: Practical Prompting

People do not dislike AI writing because a machine produced it. They dislike it because it sounds like every other piece of AI writing. The tells are consistent, learnable and — this is the useful part — removable with better instructions rather than better models.

The tells, specifically

Before you can remove them you have to be able to name them. Read any generic generated post and you will find most of these:

• The opening question. Ever wondered why…? Nobody starts a real conversation this way.
• Triads. Three adjectives, three benefits, three of everything. Human writing is lumpy; generated writing is suspiciously balanced.
• Abstraction where a detail belongs. Significant results instead of RM4,200 in the first month.
• Escalating vocabulary. Unlock, elevate, harness, leverage, seamless, game-changer, journey.
• The symmetrical close. A final line that mirrors the opening line and says nothing new.
• Uniform sentence length. Every sentence between twelve and eighteen words, forever.

Every one of these is a default the model falls back on when your prompt did not give it anything more specific to do.

Fix the input, not the output

Vague prompts produce average writing, because average is the statistically safe answer. The fix is specificity in four places: who, what belief, what evidence, and what voice.

WEAK
Write an engaging post about saving money.

WORKING
Reader: 28, salaried, take-home RM4,000, lives in KL,
thinks saving is impossible until they earn more.
Claim: an automatic RM200 transfer on payday beats budgeting apps.
Evidence: I did this for 14 months and ended with RM2,800
without ever tracking a single expense.
Voice: blunt, first person, short sentences, no motivational language.
Constraints: do not open with a question, do not use the word
"journey", no bullet points, under 120 words.

The second prompt is longer, but you write it once. In an autopilot it becomes the standing brief for every post the system generates, so the cost is paid a single time and the benefit repeats daily.

Negative constraints do the heavy lifting

Telling a model what to do produces a wide range of outputs. Telling it what not to do narrows the range immediately. Keep a personal ban list and paste it into every brief. A reasonable starting list:

Do not: open with a rhetorical question · use the words unlock,
elevate, harness, leverage, seamless, journey, game-changer ·
write in threes · end with a summary of what you just said ·
use em-dashes more than once · start sentences with "In today's".

Add to it whenever you catch a phrase repeating across your own posts. Within a few weeks the list becomes a genuine voice specification.

Feed it your own writing

The single most effective technique is also the least used: paste two or three of your own posts into the brief and instruct the model to match their rhythm. Models imitate style from examples far better than they follow adjectives. Casual means almost nothing; a sample of your actual casual writing means everything.

Ask for specificity you actually possess

A model cannot invent a true number. If your brief contains no facts, the output will be abstract — that is not a failure of the model, it is a failure of the input. So put real material in: prices, dates, timings, mistakes you made, what a customer said last week.

This is why product-driven generation works so much better than generic prompting. When a real product record with a name, a price and a description is attached, the copy has something concrete to hold on to. That is exactly what the product autopilot does — the generator follows the product's description in every post instead of guessing.

Edit like an editor, not a proofreader

Generated drafts rarely need line-level correction; they need structural cuts. Three passes, thirty seconds each:

1. Delete the first sentence. It is almost always throat-clearing. The second sentence is usually the real opening.
2. Delete the last sentence. It usually restates the post. Ending abruptly reads as confidence.
3. Replace one abstraction with one fact. Find the vaguest claim and put a number in it.

Those three edits fix the majority of generated posts, and they take less time than rewriting from scratch — which is the whole point.

Different models, different defaults

It is worth testing the same brief across two or three models before settling. They have genuinely different stylistic defaults: some are naturally terser, some more formal, some better at conversational rhythm in languages other than English. Dukun Ads lets you pick the copywriting model per autopilot and per generation in AI Studio, so running a comparison costs a fraction of a credit.

The summary you can act on

Write a brief instead of a prompt. Keep a ban list. Paste your own writing as a style sample. Put real facts in. Cut the first and last sentences. Do that and the question of whether a human or a model produced the draft stops being interesting to anyone — including your readers.

Dukun Ads stores the brief once and applies it to every scheduled post, with a test-post button so you can check the voice before anything goes live. Try it free for 3 days.