remove-ai-slop

Hamza Khalid's kit for making AI-assisted writing read as human: the Pangram flag list (words, openers, sentence shapes, punctuation, rhythm, transitions), a self-edit prompt, a chunking prompt, the 4-step StealthGPT/Pangram loop and a final seam check. Use when the user wants to humanize a draft, remove AI-sounding patterns, or get a long article past an AI detector.

SKILL.md

# Remove AI SLOP

#### The flag list, both Claude skills, and the exact chunking loop that took a 3,000-word draft from 94% AI to 100% human on Pangram.

---

🧠 **Introduction**

Most people who try to "humanize" an article do the same thing. They paste the whole draft into a humanizer, paste the result into a detector, see a red score, and try again. Same draft, same tool, same red score.

The problem was never that humanizers don't work. Pangram reads your writing closely enough that one AI-sounding paragraph can drag the whole article down, and a 3,000-word draft has dozens of places for that paragraph to hide.

Here's the shift. You stop treating the article as one piece and start working it like an assembly line. Clean the obvious; tell yourself first. Run the two skills. Cut the article into chunks. Test each chunk until it passes. Then test the seams between them.

This pack hands you that line, done. The flag list of patterns I kept cutting from flagged chunks, both Claude skills from my tests, and every prompt for the chunking loop and the final check.

**Read time:** 10 minutes. **Setup time:** about 15 minutes. **Time to first payoff:** one chunk of your own writing, passing Pangram, today.

---

### Why I Built This

Two weeks ago, I ran a 3,000-word newsletter through Pangram before sending it. It came back 94% AI.

Then I spent a week testing. The Humanizer skill barely moved a single paragraph. Adding the AntiAI skill helped, but my full draft still scored 71% AI. Every Claude-only run I tried on the full article came back above 40% AI.

What finally worked was boring: small chunks, one at a time, tested and retested, then a final check on the whole thing.

While I was doing that, I kept cutting the same things out of the chunks that failed. The same words. The same sentence shapes. The same tidy transitions. So I wrote them down.

Here's the honest part. Pangram doesn't publish a list of banned words. It's a trained model, and it judges the whole pattern, not one word at a time. So this list isn't a magic filter. It's the set of patterns that kept showing up in my flagged chunks and in well-documented signs of AI writing. Cut them first, and the tools have far less to fix.

One caveat I'll repeat inside: this fixes how your writing reads, not what it says. If the draft has no story, no opinion, and no real test behind it, a 100% human score won't make anyone want to read it.

---

✅ **Before You Start - Quick Checklist**

- [ ] You have Claude.ai or Claude Code
- [ ] You've downloaded both skills from Sections 2 and 3
- [ ] You have a StealthGPT account
- [ ] You can run Pangram scans (on pangram.com, or on your Substack draft before publishing)
- [ ] Your draft is finished and fact-checked; humanizing comes last, not first
- [ ] You have 2 or 3 paragraphs you wrote yourself, with no AI help, to use as a voice sample
- [ ] You accept that a long article is assembly-line work; some chunks will fail more than once

---

### What's Inside

**1. The Pangram Flag List:** The words, phrases, sentence structures, punctuation habits, and rhythm patterns I kept cutting from flagged chunks, each with a before and after.

**2. The Humanizer Skill:** Cleans AI-sounding words and phrases sentence by sentence.

**3. The AntiAI Skill:** Changes the rhythm so the writing stops sounding like a model.

**4. The Self-Edit Prompt:** Checks any section against the Flag List before you touch a humanizer.

**5. The Chunking Prompt:** Splits your article into 300 to 600-word chunks, one idea each.

**6. The 4-Step Loop:** The exact StealthGPT and Pangram routine, plus a tracking table.

**7. The Final Check Prompt:** Smooths the seams between chunks so the full article passes too.

---

### Who This Is For

- Newsletter writers who publish on Substack, where readers can now scan any post with Pangram
- Creators who write long-form for LinkedIn, Medium, or X and use AI in their process
- Anyone who has had their own writing flagged as AI and wants to know why
- Anyone who read the issue, tried a humanizer once, and got a red score
If you only post short updates and never check a detector, you probably don't need this. Short text is far easier to get through. This was built for long articles.

---

### How to Use It

**Step 1 - Duplicate this page:** Hit the ••• menu top right → Duplicate. Now it's yours to edit.

**Step 2 - Install both skills:** Download the files from Sections 2 and 3. In Claude Code, put each one in its own folder inside `~/.claude/skills/` (for example `~/.claude/skills/humanizer/SKILL.md`). On Claude.ai, zip each skill folder and upload it under Skills in your settings.

**Step 3 - Self-edit with the Flag List:** Section 4. Do this before any humanizer runs.

**Step 4 - Run both skills:** Sections 2 and 3, Humanizer first, then AntiAI.

**Step 5 - Chunk the article:** Section 5.

**Step 6 - Run the loop:** Section 6, one chunk at a time.

**Step 7 - Run the final check:** Section 7, then scan the full article.

🔰 **New to this?** You don't need to know how detectors work. You only need to be patient with the chunks that fail. That's a time problem, not a skill problem.

---

### 1. THE PANGRAM FLAG LIST

**What this does:** Shows you the patterns to cut before any tool touches your draft. None of these alone will get you flagged. Stacked together, they read like a model wrote the piece.

💡 **Use this before humanizing, not after.** A humanizer can only rewrite what you give it. Hand it a clean draft, and it has less to fix and fewer chances to add new tells.

#### A. Words

| Cut this | Use this instead |
| --- | --- |
| delve into | dig into, look at |
| crucial, pivotal, key (as an adjective) | important, or cut it |
| landscape (as in "the AI landscape") | name the actual thing: the tools, the market |
| showcase | show |
| underscore, highlight (as verbs) | shows, proves |
| testament | proof, or cut it |
| tapestry, vibrant, intricate | cut it, or give a specific detail |
| seamless, unlock, elevate | say what actually happens |
| foster, enhance | build, improve |
| additionally | also |
| robust, game-changer | name what makes it strong or what changed |

**Why it reads as AI:** These words show up far more often in writing produced since ChatGPT launched, and they tend to travel together. One "crucial" means nothing. "Crucial," "landscape," and "showcase" in one paragraph is a pattern.

#### B. Openers and signposts

| Cut this | Use this instead |
| --- | --- |
| "Let's dive in." / "Let's break this down." | Start with the first real point |
| "Here's the thing." / "Honestly?" | Say the thing |
| "It's important to note that..." | Just state it |
| "At its core..." / "The real question is..." | State the point plainly |
| "In today's fast-paced world..." | Cut the whole sentence |
| "Without further ado..." | Cut it |

**Why it reads as AI:** These phrases announce what's coming instead of delivering it. They make a paragraph sound like a script.

#### C. Sentence structures

**1. "It's not just X, it's Y."**
Before: It's not just a tool, it's a whole new way of writing.
After: The tool changed how I write my first drafts.

**2. Forced groups of three.**
Before: It saves time, boosts quality, and drives growth.
After: It cut my editing time in half.

**3. Sentences that end in an -ing phrase.**
Before: The detector flagged the paragraph, highlighting how fragile long-form writing is.
After: The detector flagged the paragraph. Long articles give it more to catch.

**4. "Serves as" instead of "is."**
Before: This checklist serves as your guide.
After: This checklist is your guide.

**5. Runs of short dramatic fragments.**
Before: Then it happened. No warning. No mercy. Just red.
After: The score came back red, and I had no idea why.

**6. Fake-deep formulas.**
Before: Consistency is the currency of trust.
After: Readers trust writers who show up every week.

**7. Tacked-on negations.**
Before: The prompt handles it for you, no guessing.
After: The prompt handles it, so you don't have to guess.

**Why it reads as AI:** Models reach for the same sentence shapes because they're statistically safe. Readers and detectors both learn to spot them.

#### D. Punctuation and formatting

- **Em dashes and en dashes.** Replace them with a period, a comma, a colon, or parentheses. This was the first thing I cut from every chunk.
- **Bold on every other phrase.** Keep bold for the one line that matters.
- **Lists where every item starts with a bold label and a colon.** Turn them into a sentence when you can.
- **Title Case On Every Heading.** Use normal sentence capitalization.
- **Emojis in front of every heading or bullet.**
**Why it reads as AI:** One em dash means nothing. Plenty of human writers use them. But em dashes plus bold labels plus perfect three-item lists is a pattern.

#### E. Rhythm

- **Sentences of the same length back to back.** Mix a long sentence with a short one.
- **Every paragraph built the same way.** Point, explanation, example, repeat. Break the shape on purpose.
- **A neat, upbeat ending on every section.** End on the last real fact instead.
- **A heading followed by a one-line sentence that repeats the heading.** Cut the repeat and start with the content.
**Why it reads as AI:** Human writing is uneven. AI writing tends to settle into a steady, medium-length beat.

#### F. Transitions between sections

- "Now that we've covered X, let's look at Y."
- "With that in mind..."
- "That brings us to..."
- Every section opening with a sentence that sums up the last one.
**Why it reads as AI:** These are the seams where chunks meet. Each chunk can pass on its own while the join between them still gets flagged. Section 7 fixes this.

🛡️ **What not to cut**

Some things are signs of a real person. Keep them:

- Specific, odd details only you would know
- Mixed feelings you don't fully resolve
- Asides and self-corrections in parentheses
- Short sentences and long sentences, side by side
- References tied to a specific moment, like this week's news or your own test results
Over-editing these out is how human writing ends up sounding like AI.

---

### 2. THE HUMANIZER SKILL

**What this does:** Scans text for common AI writing patterns and rewrites them, sentence by sentence. It's built on Wikipedia's "Signs of AI writing" guide.

💡 **Credit:** Humanizer is an open-source skill by Siqi Chen, released under the MIT license. The original lives at github.com/blader/humanizer. Keep the LICENSE file in the folder when you install it.

*(Not bundled here. Install Humanizer from https://github.com/blader/humanizer, MIT license.)*

```text
Use the humanizer skill on the section below.

Match my voice using this sample of my own writing:
[paste 2-3 paragraphs you wrote yourself]

Keep every fact, number, and name exactly as it is. No em dashes.
Give me only the final rewrite.

Section:
[paste section]
```

⚠️ **Honest result:** On one paragraph, this skill moved my score from 90% AI to 87% AI. On its own, it's not enough for a long article. It's the first pass, not the finish line.

---

### 3. THE ANTIAI SKILL

**What this does:** Works on rhythm instead of words. It changes sentence length and paragraph shape so the writing stops sounding like a steady AI beat.

*(Not bundled here. Download anti-ai.skill from the original kit: https://smart-scabiosa-77d.notion.site/Remove-AI-SLOP-3de816ad5f718090b554c4547cb26bdf)*

```text
Use the AntiAI skill on the section below.

Change the rhythm only. Vary sentence length, break up paragraphs
that all have the same shape, and leave lines that already sound
like me alone. Keep every fact, number, and name exactly as it is.

Section:
[paste section]
```

⚠️ **Honest result:** With both skills, my full 3,000-word draft still scored 71% AI. Better, but not a pass. That's why Sections 5 to 7 exist.

---

### 4. THE SELF-EDIT PROMPT

**What this does:** Checks any section against the Flag List and rewrites only the lines that fail, before any humanizer runs.

```text
Check the section below against this flag list:

- AI words: delve, crucial, pivotal, landscape, showcase,
underscore, highlight, testament, tapestry, vibrant, seamless,
unlock, elevate, foster, enhance, additionally, robust
- Openers and signposts: "Let's dive in", "Here's the thing",
"It's important to note", "At its core", "The real question is"
- Structures: "It's not just X, it's Y", forced groups of three,
sentences ending in an -ing phrase, "serves as" instead of "is",
runs of short dramatic fragments, tacked-on "no X" endings
- Punctuation: em dashes, en dashes, heavy bold
- Rhythm: three or more sentences in a row of similar length

List every hit with the exact line. Then rewrite only those lines.
Keep every fact, number, and name. Don't add new claims.

Match my voice using this sample:
[paste 2-3 paragraphs you wrote yourself]

Section:
[paste section]
```

---

### 5. THE CHUNKING PROMPT

**What this does:** Splits your article into chunks small enough for StealthGPT to handle and for you to test one at a time.

```text
Split the article below into chunks of 300 to 600 words.

Each chunk should cover one idea and end at a natural break.
Don't rewrite anything. Number each chunk and give it a one-line
label so I can track which ones pass Pangram.

Article:
[paste full article]
```

💡 **Why chunks:** StealthGPT has a character limit per run, so it won't take a full long article at once. Smaller chunks also make it obvious which part of the article is dragging the score down.

---

### 6. THE 4-STEP LOOP

**What this does:** Takes each chunk from flagged to passing, one at a time.

```text
1. Paste Chunk 1 into StealthGPT and run Humanize.
2. Paste the result into Pangram.
3. 100% human? Save it and move to the next chunk.
4. Any AI score? Run it through StealthGPT again and re-check.
Still failing after two runs? Split the chunk in half and run
each half on its own.
```

**Track every chunk here:**

| Chunk | Label | Run 1 score | Run 2 score | Status |
| --- | --- | --- | --- | --- |
| 1 | | | | |
| 2 | | | | |
| 3 | | | | |
| 4 | | | | |
| 5 | | | | |
| 6 | | | | |

⚠️ **Some chunks will fight you.** In my test, the first chunk passed on the first run. The second came back 55% AI and needed another pass. Dense sections with lots of ideas packed together fail most often. Split those into smaller chunks.

⚠️ **Read every output before you save it.** A rewriter can change your meaning, soften a claim, or mangle a number. A 100% human score on a wrong sentence is still a wrong sentence.

---

### 7. THE FINAL CHECK PROMPT

**What this does:** Fixes the seams. Every chunk can pass on its own while the full article still gets flagged, because the tone shifts where one chunk ends and the next begins.

```text
Here are my approved chunks, in order.

Don't change any sentence inside a chunk. Only look at the last two
sentences of each chunk and the first two sentences of the next one.
Where the tone or rhythm jumps, write one short bridge sentence in
my voice and show me exactly where it goes.

Chunks:
[paste all approved chunks in order]
```

**Then:**

```text
1. Add the bridge sentences.
2. Run each changed join through Pangram on its own.
3. Put the full article back together.
4. Scan the full article in Pangram.
5. Still flagged? Find the section it points to and send just that
chunk back through the loop in Section 6.
```

💡 **Substack writers:** You can run Pangram on your draft inside Substack before you publish. Do your final full-article scan there, since it's the same check your readers can run.

---

### Your First Article - The Run Plan

```text
Block 1 Install both skills (5 min)
Block 2 Self-edit each section with the Flag List prompt (15 min)
Block 3 Run Humanizer, then AntiAI (10 min)
Block 4 Split the article into chunks (2 min)
Block 5 Run the 4-step loop on every chunk (varies)
Block 6 Read every approved chunk for meaning (10 min)
Block 7 Run the final check prompt, add bridge sentences (5 min)
Block 8 Scan the full article in Pangram (2 min)
Block 9 Fix anything still flagged, re-scan (varies)
```

---

📋 **You're Done - Checklist**

- [ ] Both skills are installed and working
- [ ] Every section passed the Flag List self-edit before any tool touched it
- [ ] Your article is split into numbered, labeled chunks
- [ ] Every chunk scored 100% human in Pangram
- [ ] You read every chunk and confirmed nothing changed meaning
- [ ] The seams between chunks have been smoothed and re-checked
- [ ] The full article passes Pangram as one piece

---

🚫 **Common Mistakes (Avoid These)**

- [ ] Pasting the full article into StealthGPT in one go
- [ ] Running humanizers before cleaning the Flag List patterns yourself
- [ ] Testing only the chunks and never the full article
- [ ] Saving a rewritten chunk without reading it
- [ ] Cutting every em dash but leaving the same-length sentences
- [ ] Over-editing the specific, odd details that make writing sound human
- [ ] Expecting the skills alone to get a long article to 100%

---

### Closing Note

The reason your article keeps getting flagged isn't one bad word. It's that a long piece has dozens of small places where it sounds like a model, and a detector only needs a few of them.

Clean the patterns first. Work in chunks. Test the seams. The score stops being luck once the process stops changing.

**Make Yourself Irreplaceable with AI. LLMs don't think; you do.**

We tested it in public. Now go use it in private.

https://substack.com/@humzakhalid

-Hamza 💙

1 file

SKILL.md18.4 KB

--- name: remove-ai-slop description: "Hamza Khalid's kit for making AI-assisted writing read as human: the Pangram flag list (words, openers, sentence shapes, punctuation, rhythm, transitions), a self-edit prompt, a chunking prompt, the 4-step StealthGPT/Pangram loop and a final seam check. Use when the user wants to humanize a draft, remove AI-sounding patterns, or get a long article past an AI detector." metadata: author: "Hamza Khalid" source: "https://smart-scabiosa-77d.notion.site/Remove-AI-SLOP-3de816ad5f718090b554c4547cb26bdf" newsletter: "https://www.aiinpublic.com" title: "Remove AI SLOP" post: "https://www.aiinpublic.com/p/0-ai-slop-humanized-content" note: "Copied from the creator's free page. Removed only page furniture: image placeholders, empty callouts, coffee/share call-outs, lines pointing at embedded posts that did not carry over, and the closing PS. The two downloadable skills the kit links to are not bundled: Humanizer is Siqi Chen's (MIT), anti-ai.skill is linked from the original page." --- # Remove AI SLOP #### The flag list, both Claude skills, and the exact chunking loop that took a 3,000-word draft from 94% AI to 100% human on Pangram. --- 🧠 **Introduction** Most people who try to "humanize" an article do the same thing. They paste the whole draft into a humanizer, paste the result into a detector, see a red score, and try again. Same draft, same tool, same red score. The problem was never that humanizers don't work. Pangram reads your writing closely enough that one AI-sounding paragraph can drag the whole article down, and a 3,000-word draft has dozens of places for that paragraph to hide. Here's the shift. You stop treating the article as one piece and start working it like an assembly line. Clean the obvious; tell yourself first. Run the two skills. Cut the article into chunks. Test each chunk until it passes. Then test the seams between them. This pack hands you that line, done. The flag list of patterns I kept cutting from flagged chunks, both Claude skills from my tests, and every prompt for the chunking loop and the final check. **Read time:** 10 minutes. **Setup time:** about 15 minutes. **Time to first payoff:** one chunk of your own writing, passing Pangram, today. --- ### Why I Built This Two weeks ago, I ran a 3,000-word newsletter through Pangram before sending it. It came back 94% AI. Then I spent a week testing. The Humanizer skill barely moved a single paragraph. Adding the AntiAI skill helped, but my full draft still scored 71% AI. Every Claude-only run I tried on the full article came back above 40% AI. What finally worked was boring: small chunks, one at a time, tested and retested, then a final check on the whole thing. While I was doing that, I kept cutting the same things out of the chunks that failed. The same words. The same sentence shapes. The same tidy transitions. So I wrote them down. Here's the honest part. Pangram doesn't publish a list of banned words. It's a trained model, and it judges the whole pattern, not one word at a time. So this list isn't a magic filter. It's the set of patterns that kept showing up in my flagged chunks and in well-documented signs of AI writing. Cut them first, and the tools have far less to fix. One caveat I'll repeat inside: this fixes how your writing reads, not what it says. If the draft has no story, no opinion, and no real test behind it, a 100% human score won't make anyone want to read it. --- ✅ **Before You Start - Quick Checklist** - [ ] You have Claude.ai or Claude Code - [ ] You've downloaded both skills from Sections 2 and 3 - [ ] You have a StealthGPT account - [ ] You can run Pangram scans (on pangram.com, or on your Substack draft before publishing) - [ ] Your draft is finished and fact-checked; humanizing comes last, not first - [ ] You have 2 or 3 paragraphs you wrote yourself, with no AI help, to use as a voice sample - [ ] You accept that a long article is assembly-line work; some chunks will fail more than once --- ### What's Inside **1. The Pangram Flag List:** The words, phrases, sentence structures, punctuation habits, and rhythm patterns I kept cutting from flagged chunks, each with a before and after. **2. The Humanizer Skill:** Cleans AI-sounding words and phrases sentence by sentence. **3. The AntiAI Skill:** Changes the rhythm so the writing stops sounding like a model. **4. The Self-Edit Prompt:** Checks any section against the Flag List before you touch a humanizer. **5. The Chunking Prompt:** Splits your article into 300 to 600-word chunks, one idea each. **6. The 4-Step Loop:** The exact StealthGPT and Pangram routine, plus a tracking table. **7. The Final Check Prompt:** Smooths the seams between chunks so the full article passes too. --- ### Who This Is For - Newsletter writers who publish on Substack, where readers can now scan any post with Pangram - Creators who write long-form for LinkedIn, Medium, or X and use AI in their process - Anyone who has had their own writing flagged as AI and wants to know why - Anyone who read the issue, tried a humanizer once, and got a red score If you only post short updates and never check a detector, you probably don't need this. Short text is far easier to get through. This was built for long articles. --- ### How to Use It **Step 1 - Duplicate this page:** Hit the ••• menu top right → Duplicate. Now it's yours to edit. **Step 2 - Install both skills:** Download the files from Sections 2 and 3. In Claude Code, put each one in its own folder inside `~/.claude/skills/` (for example `~/.claude/skills/humanizer/SKILL.md`). On Claude.ai, zip each skill folder and upload it under Skills in your settings. **Step 3 - Self-edit with the Flag List:** Section 4. Do this before any humanizer runs. **Step 4 - Run both skills:** Sections 2 and 3, Humanizer first, then AntiAI. **Step 5 - Chunk the article:** Section 5. **Step 6 - Run the loop:** Section 6, one chunk at a time. **Step 7 - Run the final check:** Section 7, then scan the full article. 🔰 **New to this?** You don't need to know how detectors work. You only need to be patient with the chunks that fail. That's a time problem, not a skill problem. --- ### 1. THE PANGRAM FLAG LIST **What this does:** Shows you the patterns to cut before any tool touches your draft. None of these alone will get you flagged. Stacked together, they read like a model wrote the piece. 💡 **Use this before humanizing, not after.** A humanizer can only rewrite what you give it. Hand it a clean draft, and it has less to fix and fewer chances to add new tells. #### A. Words | Cut this | Use this instead | | --- | --- | | delve into | dig into, look at | | crucial, pivotal, key (as an adjective) | important, or cut it | | landscape (as in "the AI landscape") | name the actual thing: the tools, the market | | showcase | show | | underscore, highlight (as verbs) | shows, proves | | testament | proof, or cut it | | tapestry, vibrant, intricate | cut it, or give a specific detail | | seamless, unlock, elevate | say what actually happens | | foster, enhance | build, improve | | additionally | also | | robust, game-changer | name what makes it strong or what changed | **Why it reads as AI:** These words show up far more often in writing produced since ChatGPT launched, and they tend to travel together. One "crucial" means nothing. "Crucial," "landscape," and "showcase" in one paragraph is a pattern. #### B. Openers and signposts | Cut this | Use this instead | | --- | --- | | "Let's dive in." / "Let's break this down." | Start with the first real point | | "Here's the thing." / "Honestly?" | Say the thing | | "It's important to note that..." | Just state it | | "At its core..." / "The real question is..." | State the point plainly | | "In today's fast-paced world..." | Cut the whole sentence | | "Without further ado..." | Cut it | **Why it reads as AI:** These phrases announce what's coming instead of delivering it. They make a paragraph sound like a script. #### C. Sentence structures **1. "It's not just X, it's Y."** Before: It's not just a tool, it's a whole new way of writing. After: The tool changed how I write my first drafts. **2. Forced groups of three.** Before: It saves time, boosts quality, and drives growth. After: It cut my editing time in half. **3. Sentences that end in an -ing phrase.** Before: The detector flagged the paragraph, highlighting how fragile long-form writing is. After: The detector flagged the paragraph. Long articles give it more to catch. **4. "Serves as" instead of "is."** Before: This checklist serves as your guide. After: This checklist is your guide. **5. Runs of short dramatic fragments.** Before: Then it happened. No warning. No mercy. Just red. After: The score came back red, and I had no idea why. **6. Fake-deep formulas.** Before: Consistency is the currency of trust. After: Readers trust writers who show up every week. **7. Tacked-on negations.** Before: The prompt handles it for you, no guessing. After: The prompt handles it, so you don't have to guess. **Why it reads as AI:** Models reach for the same sentence shapes because they're statistically safe. Readers and detectors both learn to spot them. #### D. Punctuation and formatting - **Em dashes and en dashes.** Replace them with a period, a comma, a colon, or parentheses. This was the first thing I cut from every chunk. - **Bold on every other phrase.** Keep bold for the one line that matters. - **Lists where every item starts with a bold label and a colon.** Turn them into a sentence when you can. - **Title Case On Every Heading.** Use normal sentence capitalization. - **Emojis in front of every heading or bullet.** **Why it reads as AI:** One em dash means nothing. Plenty of human writers use them. But em dashes plus bold labels plus perfect three-item lists is a pattern. #### E. Rhythm - **Sentences of the same length back to back.** Mix a long sentence with a short one. - **Every paragraph built the same way.** Point, explanation, example, repeat. Break the shape on purpose. - **A neat, upbeat ending on every section.** End on the last real fact instead. - **A heading followed by a one-line sentence that repeats the heading.** Cut the repeat and start with the content. **Why it reads as AI:** Human writing is uneven. AI writing tends to settle into a steady, medium-length beat. #### F. Transitions between sections - "Now that we've covered X, let's look at Y." - "With that in mind..." - "That brings us to..." - Every section opening with a sentence that sums up the last one. **Why it reads as AI:** These are the seams where chunks meet. Each chunk can pass on its own while the join between them still gets flagged. Section 7 fixes this. 🛡️ **What not to cut** Some things are signs of a real person. Keep them: - Specific, odd details only you would know - Mixed feelings you don't fully resolve - Asides and self-corrections in parentheses - Short sentences and long sentences, side by side - References tied to a specific moment, like this week's news or your own test results Over-editing these out is how human writing ends up sounding like AI. --- ### 2. THE HUMANIZER SKILL **What this does:** Scans text for common AI writing patterns and rewrites them, sentence by sentence. It's built on Wikipedia's "Signs of AI writing" guide. 💡 **Credit:** Humanizer is an open-source skill by Siqi Chen, released under the MIT license. The original lives at github.com/blader/humanizer. Keep the LICENSE file in the folder when you install it. *(Not bundled here. Install Humanizer from https://github.com/blader/humanizer, MIT license.)* ```text Use the humanizer skill on the section below. Match my voice using this sample of my own writing: [paste 2-3 paragraphs you wrote yourself] Keep every fact, number, and name exactly as it is. No em dashes. Give me only the final rewrite. Section: [paste section] ``` ⚠️ **Honest result:** On one paragraph, this skill moved my score from 90% AI to 87% AI. On its own, it's not enough for a long article. It's the first pass, not the finish line. --- ### 3. THE ANTIAI SKILL **What this does:** Works on rhythm instead of words. It changes sentence length and paragraph shape so the writing stops sounding like a steady AI beat. *(Not bundled here. Download anti-ai.skill from the original kit: https://smart-scabiosa-77d.notion.site/Remove-AI-SLOP-3de816ad5f718090b554c4547cb26bdf)* ```text Use the AntiAI skill on the section below. Change the rhythm only. Vary sentence length, break up paragraphs that all have the same shape, and leave lines that already sound like me alone. Keep every fact, number, and name exactly as it is. Section: [paste section] ``` ⚠️ **Honest result:** With both skills, my full 3,000-word draft still scored 71% AI. Better, but not a pass. That's why Sections 5 to 7 exist. --- ### 4. THE SELF-EDIT PROMPT **What this does:** Checks any section against the Flag List and rewrites only the lines that fail, before any humanizer runs. ```text Check the section below against this flag list: - AI words: delve, crucial, pivotal, landscape, showcase, underscore, highlight, testament, tapestry, vibrant, seamless, unlock, elevate, foster, enhance, additionally, robust - Openers and signposts: "Let's dive in", "Here's the thing", "It's important to note", "At its core", "The real question is" - Structures: "It's not just X, it's Y", forced groups of three, sentences ending in an -ing phrase, "serves as" instead of "is", runs of short dramatic fragments, tacked-on "no X" endings - Punctuation: em dashes, en dashes, heavy bold - Rhythm: three or more sentences in a row of similar length List every hit with the exact line. Then rewrite only those lines. Keep every fact, number, and name. Don't add new claims. Match my voice using this sample: [paste 2-3 paragraphs you wrote yourself] Section: [paste section] ``` --- ### 5. THE CHUNKING PROMPT **What this does:** Splits your article into chunks small enough for StealthGPT to handle and for you to test one at a time. ```text Split the article below into chunks of 300 to 600 words. Each chunk should cover one idea and end at a natural break. Don't rewrite anything. Number each chunk and give it a one-line label so I can track which ones pass Pangram. Article: [paste full article] ``` 💡 **Why chunks:** StealthGPT has a character limit per run, so it won't take a full long article at once. Smaller chunks also make it obvious which part of the article is dragging the score down. --- ### 6. THE 4-STEP LOOP **What this does:** Takes each chunk from flagged to passing, one at a time. ```text 1. Paste Chunk 1 into StealthGPT and run Humanize. 2. Paste the result into Pangram. 3. 100% human? Save it and move to the next chunk. 4. Any AI score? Run it through StealthGPT again and re-check. Still failing after two runs? Split the chunk in half and run each half on its own. ``` **Track every chunk here:** | Chunk | Label | Run 1 score | Run 2 score | Status | | --- | --- | --- | --- | --- | | 1 | | | | | | 2 | | | | | | 3 | | | | | | 4 | | | | | | 5 | | | | | | 6 | | | | | ⚠️ **Some chunks will fight you.** In my test, the first chunk passed on the first run. The second came back 55% AI and needed another pass. Dense sections with lots of ideas packed together fail most often. Split those into smaller chunks. ⚠️ **Read every output before you save it.** A rewriter can change your meaning, soften a claim, or mangle a number. A 100% human score on a wrong sentence is still a wrong sentence. --- ### 7. THE FINAL CHECK PROMPT **What this does:** Fixes the seams. Every chunk can pass on its own while the full article still gets flagged, because the tone shifts where one chunk ends and the next begins. ```text Here are my approved chunks, in order. Don't change any sentence inside a chunk. Only look at the last two sentences of each chunk and the first two sentences of the next one. Where the tone or rhythm jumps, write one short bridge sentence in my voice and show me exactly where it goes. Chunks: [paste all approved chunks in order] ``` **Then:** ```text 1. Add the bridge sentences. 2. Run each changed join through Pangram on its own. 3. Put the full article back together. 4. Scan the full article in Pangram. 5. Still flagged? Find the section it points to and send just that chunk back through the loop in Section 6. ``` 💡 **Substack writers:** You can run Pangram on your draft inside Substack before you publish. Do your final full-article scan there, since it's the same check your readers can run. --- ### Your First Article - The Run Plan ```text Block 1 Install both skills (5 min) Block 2 Self-edit each section with the Flag List prompt (15 min) Block 3 Run Humanizer, then AntiAI (10 min) Block 4 Split the article into chunks (2 min) Block 5 Run the 4-step loop on every chunk (varies) Block 6 Read every approved chunk for meaning (10 min) Block 7 Run the final check prompt, add bridge sentences (5 min) Block 8 Scan the full article in Pangram (2 min) Block 9 Fix anything still flagged, re-scan (varies) ``` --- 📋 **You're Done - Checklist** - [ ] Both skills are installed and working - [ ] Every section passed the Flag List self-edit before any tool touched it - [ ] Your article is split into numbered, labeled chunks - [ ] Every chunk scored 100% human in Pangram - [ ] You read every chunk and confirmed nothing changed meaning - [ ] The seams between chunks have been smoothed and re-checked - [ ] The full article passes Pangram as one piece --- 🚫 **Common Mistakes (Avoid These)** - [ ] Pasting the full article into StealthGPT in one go - [ ] Running humanizers before cleaning the Flag List patterns yourself - [ ] Testing only the chunks and never the full article - [ ] Saving a rewritten chunk without reading it - [ ] Cutting every em dash but leaving the same-length sentences - [ ] Over-editing the specific, odd details that make writing sound human - [ ] Expecting the skills alone to get a long article to 100% --- ### Closing Note The reason your article keeps getting flagged isn't one bad word. It's that a long piece has dozens of small places where it sounds like a model, and a detector only needs a few of them. Clean the patterns first. Work in chunks. Test the seams. The score stops being luck once the process stops changing. **Make Yourself Irreplaceable with AI. LLMs don't think; you do.** We tested it in public. Now go use it in private. https://substack.com/@humzakhalid -Hamza 💙