Make ChatGPT Write Like a Human: The Anti-AI-Tell System Prompt

If you've been editing AI drafts for more than a few weeks, you've developed an ear for it. The hedge at the top of every paragraph. The triplet that shows up whether or not three things actually exist. The closing sentence that restates everything you just read, as if the model is nervous you didn't catch it.

That voice is a liability now. Google's helpful content guidance, AdSense reviewers, and - most importantly - real readers can all hear it. The fix isn't to stop using AI. The fix is to stop shipping AI's default voice.

Here's the system prompt I run before every draft, the seven manual edits that catch what the prompt misses, a before/after example, and a checklist you can print out and keep next to your screen.

What "AI Voice" Actually Sounds Like

If you can't see the pattern, you can't fix it. Here are the seven tells that mark almost every default ChatGPT or Claude output:

Hedged openers. "In today's rapidly evolving landscape...", "When it comes to...", "It's no secret that..."
Tidy summary closers. Every paragraph ends with a one-sentence wrap-up that restates what you just read.
Triplets where doubles or singles work. "Strategic, scalable, and sustainable." "Plan, execute, and refine." Reflexive triadic phrasing.
Hedge words. "Often," "typically," "may," "could," "various," "numerous." Used to avoid making any concrete claim.
Corporate verbs. "Leverage," "navigate," "unlock," "harness," "drive," "facilitate." Almost never the right word.
Bullet bloat. Six-point lists where two points would carry the meaning. The model loves a list.
Faux humility prefaces. "It's worth noting that...", "While there are exceptions...", "Important to mention..." Throat-clearing before the actual point.

You'll spot these everywhere now. You can't unsee them.

Before and After: Seeing the Transformation

This is what a raw AI paragraph typically looks like before any editing pass:

Before (raw AI output): "In today's fast-paced digital landscape, it's no secret that content creators are increasingly leveraging AI tools to navigate the complexities of modern content production. While there are certainly many approaches one might take, it's worth noting that the most effective strategies often involve a combination of careful prompt engineering, thoughtful editing, and a commitment to maintaining authentic voice. By implementing these key principles, solopreneurs can harness the full potential of AI-assisted writing."

Every single tell is present. The opener. The hedge words. "Leverage." "Navigate." "Harness." The closing restatement. This is what ships when you trust the default.

Here's the same paragraph after one editing pass using the system prompt below and the seven-edit checklist:

After (one editing pass): "AI tools can produce a full draft in 30 seconds. The problem is that 30-second draft reads like a press release. This guide gives you the exact system prompt and edit sequence that strips out the corporate filler - so the output sounds like you wrote it, not a committee."

Shorter. Concrete. No hedge words. The reader gets a specific promise instead of a vague claim about "harnessing potential."

I edit 15-20 AI drafts a week for this site. The before above is not exaggerated - it's nearly copy-pasted from a real output I got last Tuesday. The after took about 90 seconds.

- Pete Fluriach, Miami

The Anti-AI-Tell System Prompt

Drop this into the System Prompt field of any custom GPT, Claude project, or Gemini gem:

You are a sharp, experienced writer who writes the way smart people
talk. You write for adults. You operate under these rules:

VOICE
- Write the way you'd explain it to a smart colleague over coffee
- Use one-syllable words when they fit
- Use sentence fragments when they punch better. Like this.
- Cut every hedge word: often, typically, may, could, various, numerous
- Cut every corporate verb: leverage, navigate, unlock, harness, drive, facilitate

STRUCTURE
- Open with the most specific, concrete sentence you can. No throat-clearing
- Don't end paragraphs with summary sentences. Let the last point land
- Avoid triplets unless three things actually exist. Two or four also work
- Default to fewer bullet points, not more. Lists earn their slots

CLAIMS
- Make concrete claims, not safe generalizations
- If you'd write "X can be very impactful," instead write what specifically X does
- Numbers and proper nouns beat adjectives

NEVER
- "In today's fast-paced/rapidly-evolving/digital world"
- "It's no secret that...", "It goes without saying...", "It's worth noting..."
- "Whether you're a [X], [Y], or [Z]..."
- "By following these tips...", "In conclusion...", "To summarize..."
- "Game-changer," "deep dive," "level up," "unlock," "supercharge"
- Em-dashes for stylistic flourish - only when grammatically required

That last line is the cheat code. ChatGPT is famously addicted to em-dashes. Banning the stylistic ones immediately strips one of the most detectable tells.

The system prompt won't catch everything. It will catch most of it. After it ships output, you run the manual pass.

Three Copy-Pasteable Prompts for Specific Situations

The system prompt above handles new drafts. These three prompts handle the other common situations: rewriting existing AI content, torching a specific set of tells, and matching a voice you've already built.

Prompt 1: Rewrite in my voice

Rewrite the following passage to match my editorial voice. Rules:
- Shorter sentences. Average under 18 words.
- Cut every hedge: often, typically, may, could, various, numerous
- Cut every corporate verb: leverage, navigate, unlock, harness
- No triplets unless three distinct things exist
- No summary sentences at paragraph ends
- Open with a specific claim or action, not a generalization
- Sound like a smart person talking, not a document

[paste your draft here]

Prompt 2: Kill the AI tells checklist pass

Audit this draft for AI writing tells and rewrite each flagged instance.
Check for and fix every one of the following:

1. Hedged opener ("In today's world...", "It's no secret...", "When it comes to...")
2. Summary closer (last sentence of paragraph restates what came before)
3. Reflexive triplet (three items where one or two would work)
4. Hedge word (often, typically, may, could, various, numerous, generally)
5. Corporate verb (leverage, navigate, unlock, harness, drive, facilitate)
6. Negative parallelism ("It's not just X, it's Y" - flag and rewrite as a direct claim)
7. Faux humility preface ("It's worth noting...", "Important to mention...", "While there are exceptions...")
8. Em-dash used for stylistic flourish (replace with period, comma, or colon)

Return the revised draft with a short note for each fix you made.

[paste your draft here]

Prompt 3: Match a writing sample

I'm going to give you two things: a writing sample that represents my voice,
and a draft that needs rewriting to match that voice.

Analyze the sample for: sentence length, punctuation habits, level of formality,
use of fragments, and tone. Then rewrite the draft to match those patterns.
Do not add hedges, summaries, or transitions the sample doesn't use.

MY VOICE SAMPLE:
[paste 2-3 paragraphs of your own writing]

DRAFT TO REWRITE:
[paste the AI draft]

I use Prompt 2 on about half of the drafts I edit. When I can't place what's making a piece feel "off," running the checklist pass surfaces it within seconds - usually it's the negative parallelism or the closing summary sentences I missed on the first read.

- Pete Fluriach

The AI Writing Tell Checklist

Print this. Keep it next to your screen for the first few weeks. Once you internalize the patterns, you won't need it.

[ ] Hedged opener - first sentence starts with "In today's...", "It's no secret...", "When it comes to...", or any broad scene-setting claim
[ ] Summary closer - last sentence of a paragraph restates the paragraph's main point instead of advancing it
[ ] Reflexive triplet - three-item list or phrase where one or two items would carry the meaning
[ ] Hedge word - "often," "typically," "may," "could," "various," "numerous," "generally," "sometimes" used to soften a claim the model should make directly
[ ] Corporate verb - "leverage," "navigate," "unlock," "harness," "drive," "facilitate," "empower," "optimize" as a substitute for a specific verb
[ ] Negative parallelism - "It's not just X, it's Y" or "Not only X, but also Y" used as a rhetorical flourish rather than a genuine contrast
[ ] Faux humility preface - "It's worth noting that...", "Important to mention...", "While there are certainly exceptions..." before the actual point
[ ] Vague attribution - "Many experts agree...", "Studies show...", "Research suggests..." with no specific source
[ ] Inflated transition - "Building on that foundation...", "With that in mind...", "Taking it a step further..." between paragraphs that don't actually build on each other
[ ] Em-dash overuse - more than one em-dash per 300 words, or em-dashes used where a period or comma would work

Run through this list top-to-bottom on any AI draft before it publishes. Most drafts will have 4-6 hits. Getting to zero on every item takes about 10 minutes. That 10 minutes is the difference between content that reads like content and content that reads like yours.

The Seven-Edit Manual Pass

These are the find-and-replace passes I run on every AI draft before it leaves my desk. Each takes about 20 seconds.

1. Cut the first sentence

Almost every AI paragraph opens with throat-clearing. Cut the first sentence and read again. About 60% of the time, the paragraph is sharper without it.

2. Replace every "leverage" with a real verb

leverage - use, rely on, run on, build with, plug into. Pick the verb that's actually true.

3. Hunt the hedge words

Search for: often, typically, may, could, various, numerous, generally, sometimes. For each one, ask: is the hedge true, or is the model softening a claim it should make confidently? Most are softeners. Cut them.

4. Kill the closing summary sentences

Read each paragraph's last sentence. If it summarizes what came before instead of advancing the argument, delete it. Let the previous sentence land.

5. Break a long sentence with a fragment

AI tends to write 25-word sentences with three commas. Find one. Break it into a regular sentence and a 4-word fragment that punches.

Before: "While most solopreneurs initially struggle with prompt engineering,
which can feel overwhelming at first, those who persist find that the skill
compounds quickly."

After: "Most solopreneurs struggle with prompt engineering at first.
The skill compounds fast."

Better.

6. Find the bullet list and cut a third of it

Default ChatGPT lists run 5-7 items. The first 2-3 are the real ones; the rest are filler. Cut the bottom third.

7. End on a specific, concrete sentence

The last sentence of a piece is where AI is weakest. It loves a generic platitude - "Embrace the journey." "Start small and stay consistent." Replace with a specific action. "Open ChatGPT, paste this prompt, change one variable." Concrete instructions outperform inspiration every time.

What Survives

Some "AI tells" are actually just good writing. Don't strip these:

Numbered steps when the order matters. That's not AI voice - that's clarity.
Definitions before unfamiliar terms. Helping your reader is good writing.
Headers that promise specifics. "Three Edits to Try Today" beats "Some Thoughts on Editing."

The goal isn't to make your writing sound like nobody used AI. The goal is to make it sound like you. AI is the first draft engine. You're the editor. The system prompt above stops the model from imposing its default voice. The seven edits replace it with yours.

Open ChatGPT right now. Paste the system prompt. Run it on the last thing you published. Then run the checklist. Count how many hits you get. If it's more than three, the 10-minute editing pass was worth learning.