How to Write a System Prompt That Strips Out AI Tells

If you paste the same generic "write naturally" instruction into ChatGPT and wonder why the output still sounds like a corporate brochure, this is why: vague instructions produce vague results. A system prompt that actually strips AI tells has to be specific about the problems it's fixing.

This guide shows you how to write one from scratch (or sharpen the one you already have) so the model stops defaulting to its most confident, most robotic register.

Why system prompts work better than one-off instructions

When you type "sound more human" in the chat, the model treats it as a soft suggestion. It might dial back the bullet points slightly, then slip right back into "Furthermore, it is important to note" in the next paragraph.

A system prompt is different. It sits above the conversation and shapes every response from the start. The model reads it before it reads your request, so the stylistic constraints get baked into the generation rather than applied as an afterthought.

That said, a system prompt is only as good as what you put in it. The goal is to be prescriptive without being so long that the model starts cherry-picking which rules to follow.

The specific habits you need to ban by name

AI models have default patterns that show up across ChatGPT, Claude, and Gemini because they all trained on similar corpora. If you don't call these out explicitly, the model won't suppress them.

Sentence openers that flag AI instantly

A few openers appear constantly in AI output and almost never in natural writing:

• "Certainly!"
• "Great question!"
• "It's important to note that..."
• "In today's fast-paced world..."
• "This comprehensive guide will..."

Add a line to your system prompt that lists these verbatim and marks them off-limits. "Do not begin a response with affirmations like 'Certainly,' 'Of course,' or 'Great question'" is clear enough that the model can follow it.

Vocabulary tells

There's a cluster of words that AI models reach for when they want to sound impressive. You know most of them if you've spent time editing AI copy: delve, tapestry, testament, seamless, robust, pivotal, leverage (used as a verb), underscore (also as a verb), foster, garner, navigate (when talking about abstract challenges), and elevate.

None of these are wrong in isolation, but they appear at statistically unusual rates in AI-generated text. List them in your system prompt under a heading like "Banned vocabulary" and tell the model to use simpler, more direct alternatives.

The em-dash problem

Models love the em dash. A piece of AI writing will often have six or eight of them in a single article, even in paragraphs where a period or comma would read more naturally. Limit em dashes to two per piece, or ban them outright and tell the model to prefer commas, parentheses, or a new sentence.

Structural over-reliance

AI defaults to three-part lists ("First... Second... Third..."), symmetrical parallel structures, and what editors call "the rule of three." Real prose doesn't have this compulsion toward tidy groupings. A sentence that says "It saves time, reduces errors, and boosts confidence" is not necessarily better than one that just says "It's faster and less error-prone." Tell the model to resist three-part structures when two or four would be more honest.

How to write the voice section

Banning bad habits is only half the job. You also need to tell the model what voice to aim for. This is where most system prompts fall apart because they're too abstract: "Write conversationally" means nothing specific to a language model.

Concrete alternatives:

Sentence length variation. Tell the model to mix short sentences with longer ones. Three long sentences followed by one very short one creates a rhythm that feels human. Give a target ratio: "Aim for sentences between 8 and 25 words, with occasional sentences under 8 words for emphasis."

Real point of view. AI hedges constantly. "Some experts argue..." and "It is worth considering..." are deflections. If you want the model to sound like a person, tell it to take positions: "State opinions directly. Do not hedge every claim with 'some argue' or 'it's worth noting.'"

Specific examples over abstract claims. AI often writes at altitude ("This approach drives engagement and supports growth") because it's safer than committing to specifics. A good system prompt tells the model to ground abstract claims in concrete examples or data points rather than restating them at a higher level of generality.

Before and after: what the prompt actually changes

Here's what a paragraph looks like without a humanizing system prompt:

In today's fast-paced digital landscape, it is crucial to leverage seamless content strategies that foster engagement and drive robust results. This comprehensive approach underscores the importance of authenticity in modern communication.

And here's what the same request produces when the system prompt bans the vocabulary, the openers, and the empty abstractions:

Readers can tell when something was written by a machine. The giveaways are usually vocabulary (words like "leverage" and "foster" in the same sentence) and structure (three parallel clauses, every time). Fixing both in your system prompt makes a measurable difference in how the output reads.

The second version isn't perfect prose. It's a model output, not a human draft. But it doesn't trigger the same instant recognition. That's the realistic goal of a system prompt: fewer flags, more plausible deniability, less editing work on your end.

You can experiment with a ready-made version at /humanizer-prompt. It's a starting point you can copy and modify rather than building from zero.

How long should a system prompt be?

There's a practical upper limit. Beyond roughly 400 to 500 words, models start treating parts of the system prompt as background noise, especially on long conversations. The most important constraints get buried.

Keep the core rules short and specific:

1. A banned-vocabulary list (10 to 20 words)
2. A banned-opener list (5 to 10 phrases)
3. One or two voice instructions (sentence length, hedging policy)
4. A structural rule (limit em dashes, avoid rule-of-three)

That's usually enough to see a clear difference. If you need the model to match a specific person's voice or a specific brand's tone, add a short example paragraph. "Write in a voice similar to this:" followed by 3 to 4 sentences of target prose works better than any abstract description.

For a deeper comparison of how this plays out across different models, see humanizer prompts for ChatGPT, Claude, and Gemini compared.

Iterating on your prompt

The first version you write won't be perfect. Plan to test it against a few requests and look for what slips through.

Common gaps after a first pass:

• The model drops the banned vocabulary but keeps the parallel structure
• Sentence length varies within a paragraph but not between paragraphs
• Hedging language disappears but filler transitions ("It is also worth mentioning...") survive

Each time you notice a pattern, add a rule. After two or three iterations, most people find the prompt stabilizes. The same patterns stop recurring.

The best prompt to make AI writing sound human has a version you can use as a base for iteration, already tested across common use cases.

One more thing worth doing: read the output aloud. If you stumble anywhere, that's usually where the AI register crept back in. Your ear catches what a checklist misses.

---

FAQ

Does a system prompt work the same way as custom instructions in ChatGPT?

Custom instructions in ChatGPT are a form of persistent system prompt. They apply to every conversation automatically. A system prompt you set through the API or via a tool that lets you configure it per-session works the same way mechanically, but it only applies to that specific conversation. For most people, custom instructions are the practical option because they don't require API access.

Will a humanizing system prompt guarantee my content passes AI detectors?

No, and be skeptical of anyone who claims otherwise. Detection tools look at statistical patterns in text, and those patterns change as models update. A well-written system prompt reduces common tells, but it doesn't produce guaranteed detector-proof output. The goal is copy that reads naturally to humans, which is a different (and more meaningful) target than gaming a detection algorithm. For more on how detection actually works, see a ChatGPT prompt that makes copy sound like you wrote it.

How specific do I need to be about the voice I want?

Very specific, or not at all. Vague instructions like "write conversationally" produce inconsistent results. Either give concrete rules ("sentences should vary between 8 and 30 words; take direct positions; avoid hedges") or give an example paragraph as a style reference. Something vague like "write like a knowledgeable friend" leaves the model filling in the gaps with its own defaults, which you probably won't like.

Can I use the same system prompt across ChatGPT, Claude, and Gemini?

Mostly yes, with minor adjustments. The banned vocabulary and structural rules transfer cleanly because the tells are similar across models. Where models differ is in how they handle direct instructions versus example-based guidance. Claude tends to respond well to explicit rules; ChatGPT-4 and Gemini often do better with a target-prose example in addition to the rules. Test your prompt on each platform and note where the gaps are.

What's the single most impactful rule I can add right now?

Ban the opener phrases. "Certainly," "Great question," "Of course," and "It's important to note" are the clearest, most universal AI tells. Adding "Do not begin any response with an affirmation or filler phrase" costs you four seconds to type and eliminates some of the most recognizable AI patterns in one rule.