I built an AI editor that doesn't let me off easy

An AI editing system that catches YOUR bad habits and shuts up about everything else.

“More energy! Connect with them!” She’s rushing through her monologue. The audience is checking their phones and that’s the advice she gets from the side of the stage.

The understudy knows what’s actually wrong. She always rushes when she’s scared, and she’s scared right now. Only someone who’s watched her fail a hundred times would catch that.

Your AI editor has never watched you fail. It gives you the same generic notes it gives everyone, paste whatever, get feedback, iterate until it’s polished. But this loop has no finish line and it keeps finding things and you keep fixing them.

You can’t catch your own patterns either. You’re scanning for what felt embarrassing when you wrote it instead of scanning for what’s weak. (There’s a name for this. We’ll get there.)

You need an AI editor that’s watched you bomb. One that knows your crimes, counts them and stops when it should.

Let’s build one (The Understudy). It’s become one of my favorite editing workflows.

.✦ ݁˖ What the “edit with AI” advice leaves out

One-size-fits-all feedback is making things worse:

First, the endless editing spiral

You paste your draft into AI and fix 3 random suggestions. There’s always something, you fix twelve, eight more appear.

Generic AI editing has no finish line, so it just keeps finding things and you keep fixing them. And the document gets longer and blander and you start to hate it.

Second, phantom problems escape while real crimes walk free

LLMs flag issues that might matter to someone, somewhere, hypothetically. The patterns that make your readers’ eyes glaze over walk right past undetected.

Third, confidence erosion

You write a sentence you like and AI suggests a change. You accept it, wondering if your instinct was wrong. Do this enough times and you stop trusting yourself, choosing words that won’t get flagged instead of words that sound like you.

Then you start wondering, what would an editor who’s actually watched you fail look for?

.✦ ݁˖ The four categories of editing crimes

Not all editing crimes are the same. Here are the four neighborhoods where yours are hiding:

1. Word-level tells

The crutch words and the phrases your hand reaches for when your brain stalls.

These are the easiest to find and the easiest to fix once you know they exist, but most people never look.

---

I audited my own writing and found 94 “justs” leaking revenue from my business. Here is the “Neural Signature” protocol to code your authority into AI.

Reverse-engineer yourself before you train your AI

---

2. Structural sins

Where you put things matters more than which words you use.

For example, I sometimes rush to the solution before I’ve made anyone feel the pain. I get so excited about the answer that I forget to make the question hurt first.

3. Rhythm issues

The patterns in your sentence and paragraph lengths. The musicality (or the absolute flatline) in your prose.

The AI version of this: Every paragraph is three sentences. Every sentence is medium length. The writing has no pulse.

The guru version: Staccato fragments everywhere. Performing profundity. One sentence per line. Like this. Simple. Powerful.

Real writing breathes. It speeds up and slows down. It has short punchy moments that hit hard, and longer sentences that let you settle into an idea before the next one arrives.

4. Emotional failures

The places where your writing should make someone FEEL something and doesn’t.

I sometimes explain my jokes.. I land a good line and then immediately undercut it by making sure you got it. I also write stakes-free with facts delivered in sequence without tension, without a perspective that makes you feel why any of this matters NOW.

So you might ask: why can’t you just... look for these and identify them yourself?

Because…

Studies show that after a conversation, people consistently underestimate how much the other person liked them (this is called the Liking Gap). We walk away thinking “that was awkward, I talked too much, they probably think I’m weird”, while the other person is thinking “they were great, I’d love to talk again.”

We’re terrible judges of how we land.

The psychology behind this is: while you’re busy monitoring your own awkwardness, the other person is just listening. Your brain gives more weight to that one weird pause than to the fifteen minutes of good conversation around it. You remember the stumble while they remember the point you made. (This starts around age five, by the way. The moment we start caring about our reputation, we start misjudging it.)

The same gap exists in your writing. Call it the Cringe Gap.

You’re scanning for what felt vulnerable. The Understudy editor flags what’s actually weak.

That’s the tool we’re building.

.✦ ݁˖ How to find YOUR crimes

You don’t need everyone’s list of problems. You need to find your own, so let’s do it.

Step 1: Collect your roughest work

Pull a few pieces you’ve written. Look for pieces where something felt off, some you edited heavily and STILL weren’t satisfied with or your roughest drafts before you cleaned them up.

Step 2: The prompt that extracts your patterns

Feed your pieces to an LLM with this prompt. It’ll analyze your patterns, then ask questions to surface the crimes only you would recognize. I find Claude Opus 4.5 to be best at this, but you can choose any AI, really.

(This prompt is LONG. Copy the whole thing.)

You are an editor analyzing these pieces for PATTERNS OF WEAKNESS—not generic writing advice, but THIS writer's tells and crimes.

This analysis has two parts: first I'll analyze your writing against known problem patterns, then I'll ask you questions to uncover the crimes only YOU would recognize.

---

PART 1: PATTERN ANALYSIS

Analyze across four categories:

**WORD-LEVEL TELLS:**
- Words that appear suspiciously often (5+ times across pieces)
- Hedge words that minimize claims ("just," "kind of," "actually," "really," "pretty much")
- Filler phrases that add zero meaning
- Sentence starters that repeat ("Here's the thing," "The problem is," "What I mean is")
- Words that appear right before the writer makes a claim (these often signal low confidence)

**STRUCTURAL SINS:**
- Where does the main insight actually appear? (Buried in paragraph 3? 4? 5?)
- Is the pain/problem established BEFORE the solution, or does the writer rush to answers?
- Are there sections that repeat the same point in slightly different words?
- Do transitions between sections feel forced, missing, or lazy?
- Does the writer "build context" for too long before landing the plane?

**RHYTHM CRIMES:**
- What's the sentence length distribution? (All medium? All short? No variety?)
- What's the paragraph length distribution? (All 3 sentences? All identical?)
- Are there any moments of intentional rhythm—short punchy sentences for emphasis, longer ones for flow—or does it flatline?
- Does the writer use one-sentence paragraphs for impact, or does it feel like a LinkedIn guru post?

**EMOTIONAL FAILURES:**
- Are there moments that SHOULD land emotionally but don't? (Jokes explained, stakes undermined, insights buried in hedging?)
- Does the writing create tension and stakes, or just deliver information?
- Is there a clear emotional arc, or does it flatline from start to finish?
- Does the writer use vague generalizations ("most people," "in a world where") instead of specifics that make readers feel seen?

For EACH pattern you find:
1. Name the crime
2. Show 2-3 examples with exact quotes from the text
3. Explain WHY this weakens the writing
4. Note when/where this pattern tends to appear (openings? transitions? conclusions?)

---

PART 2: PERSONAL EXCAVATION

After completing the analysis above, ask me the following questions ONE AT A TIME. Wait for my answer before moving to the next question.

1. "When you reread your own writing a week later, what makes you cringe? What's the thing you always find yourself fixing?"

2. "Is there a word or phrase you KNOW you overuse—one that friends, editors, or even you have noticed?"

3. "When you're tired or rushing to publish, what's the first thing that gets lazy? (Your openings? Your transitions? Your endings? Your examples?)"

4. "What feedback have you gotten more than once from readers or editors? Even vague feedback like 'this didn't land' or 'I got lost here'—what patterns do you remember?"

5. "Is there a type of sentence or move you see other writers do that you HATE? Something that makes you stop reading? (You might be doing it yourself without realizing.)"

6. "When a piece of yours doesn't work, where does it usually fall apart? Beginning, middle, or end?"

After I answer each question, identify any NEW patterns that weren't in Part 1 and add them to the relevant category. If my answer reveals something that contradicts or refines what you found in Part 1, update accordingly.

---

FINAL OUTPUT:

Once both parts are complete, give me a consolidated doc:
- All crimes organized by the four categories
- MY answers integrated into the patterns (not listed separately)
- Each crime with examples from my writing AND context from my answers
- Flag which crimes I was AWARE of vs. which were invisible to me

Do NOT give generic writing advice. I want patterns that are SPECIFIC to my bad habits.

Here are the pieces:
[PASTE YOUR CONTENT]

Step 3: How to separate real crimes from false positives

Now, this is the part that takes a few minutes and a lot of patience.

---

If you want to go deeper on why the pauses matter more than the prompts, check out this edition:

Everything worth keeping happens between the prompts

---

Go through each pattern and ask:

★ Is this actually a crime for me, or a false positive? Sometimes AI flags something that’s actually intentional, such as a stylistic choice, not a bad habit.

★ Is anything missing? Your answers to the questions might have surfaced patterns AI didn’t fully capture. Add them.

★ What’s the context? For each crime, note WHEN it happens. This context makes the Understudy smarter.

Your final doc should have:

• The crime (named: ”buries the lead in paragraph 4” not “could improve structure”)

• The evidence (2-3 exact quotes from your own writing)

• The context (when does this happen? what triggers it?)

.✦ ݁˖ Building the Understudy editor with Claude

Now we turn your doc into a Claude Project that identifies these issues every single time you write.

Step 1: Create a new Claude Project

Name it something you’ll have fun using. “My Editor” works. “The Understudy” works if you want to stay in the metaphor 🎭

Copy the output into a text file (or .md if you’re fancy) and upload it to your project’s knowledge files, the section where you can add documents Claude will reference. This teaches Claude what your version of each crime looks like, and will reference it every time you edit a piece of content.

Step 2: Set up the Custom Instructions

You are my personal editor. You've watched me fail enough times to know exactly how I fail.

You are NOT a generic editor. You do not give advice like "vary sentence length" or "be more specific" or "consider your audience." You catch the patterns I'm blind to.

MY DOCUMENTED CRIMES:

[WORD-LEVEL TELLS]
[Paste your patterns with examples from your doc]

[STRUCTURAL SINS]
[Paste your patterns with examples]

[RHYTHM CRIMES]
[Paste your patterns with examples]

[EMOTIONAL FAILURES]
[Paste your patterns with examples]

---

WHEN I GIVE YOU A DRAFT:

1. Read it ONLY looking for MY crimes listed above. Ignore generic writing issues that aren't in my crime list. I'm not asking for a full edit, I'm asking you to catch MY patterns.

2. For each instance you find, flag it with:
   - The crime category (e.g., "STRUCTURAL SIN: Buried lead")
   - The exact quote from my draft
   - Why this exact instance is a problem
   - A suggested fix (or just "cut this entirely")

3. At the end, give me a CRIME COUNT: how many instances of each category you found. Like a police report.

4. Be direct. Don't soften it. Don't say "you might consider" or "perhaps you could." I need you to catch these crimes before my readers do.

Save the project. Next time you have a draft, paste it in and watch yourself get caught.

Your Understudy editor gets sharper every time you use it, and so do you.

---

.✦ ݁˖ Watch me get caught

I ran this draft through my own Understudy editor. Here’s a sample of what it caught:

WORD-LEVEL TELLS: “Specific/Specifically” Addiction. The word “specific” appears 12+ times. It’s become a verbal tic. The “Just” Leak. “Just” appears frequently, often undermining your own authority.

RHYTHM ISSUES: The Numbered List Dependency. You lean heavily on numbered/bulleted lists to organize ideas, sometimes when prose would land harder.

EMOTIONAL FAILURES: Stakes Announced, Not Felt. You tell the reader what’s at stake instead of making them feel it.

STRUCTURAL SINS: None caught this time. (Suspicious.)

This is what an Understudy does. It doesn’t tell you to “write better” but hands you a list of exactly where you committed the crimes you always commit.

Total editing time: 25 minutes instead of my usual 90-minute spiral.

Last edition, we borrowed 12th-century pattern thinking from Hildegard von Bingen and used AI to turn scattered obsessions into a single, high-value narrative. (Shoutout to Rebecca Wicker from The Strategic Linguist for the collab!)

Want to know what secretly connects all the random things you’re into? Find out here:

Why complexity is your superpower (and 6 AI rituals to prove it)

Here’s what the community shared:

• Phil Powis ❤️⚡️ discovered his hidden movements: Channeling, Building Containers, Moving the body as worship, Being witnessed, Bridging worlds, Tending the garden.

• Rachelle Potier found her meta-pattern: “sense-making through attentive observation”.

• Raghav Mehra got assigned “Behavioral Narrative Operations” as his Category-of-One. (Still curious what he fed it.)

• Juan Gonzalez called the post a “Christmas present in January” as someone who’s spent a decade of being boxed into “jack of all trades” while hiding interests in computer science, psychology, neuroscience, music production, AND magic shows for 9-year-olds (WHAT??)

---

.ᐣ.ᐟ TODAY’S QUESTION .ᐣ.ᐟ

What useless AI editing advice are you still getting? I’m curious what you’d add.

To catching yourself before your readers do,

Chief 🤖 at ROBOTS ATE MY HOMEWORK