How to Make AI Write Like You, Not a Robot

Ellie at Willa, Editorial lead, Willa · Published August 12, 2026 · Updated August 25, 2026

You can spot it instantly: the sycophantic opener, the "it's worth noting," the em dashes marching through every paragraph. AI-generated writing has a look, and once you see it, you can't unsee it. On LinkedIn, in blog posts, in your own inbox. At Willa's live session "Em Dash Rehab," Maddie Engelmeier, who spent five years at Motive and led its AI transformation over her final year and a half, walked members through the system she built to fix it: six files that teach AI to write in her actual voice, then scrub out the telltale patterns before she hits send. "The AI slop was being generated by me, along with everybody else, and it was annoying," she admitted. "I found myself actually rewriting it, spending more time proofreading it and editing it, or just not using it at all." Here's how she solved it.

Why does AI writing all sound the same?

Before you can fix AI slop, you have to name it. Maddie opened by dissecting a typical AI-generated email; the tells fall into a few buckets:

• The sycophantic opener and performative empathy: pretending to care, over the top, with nothing specific attached.
• Over-signposting ("it's worth noting"), false closes, and the mechanical opener-body-closer formula applied to every message regardless of context.
• Signature words (delve, tapestry, multifaceted, nuanced, robust, "at the intersection of") and over-coupled adjective-noun pairs like "significant opportunity."
• Structural tics: the tricolon reflex ("it's fast, flexible, and scalable"), uniform sentence lengths, overly smooth transitions, the concluding summary nobody asked for.
• And yes, em dash overuse. As Maddie pointed out, though, punctuation is the easiest tell to fix. The harder ones are structural.

Humans don't format sentences that crisply across the board, and they don't use false energy where it's not needed. Recognizing these patterns in your own AI-assisted writing is step one. Maddie was candid that six months earlier, she was producing them too.

Why doesn't "make it sound like me" prompting work?

Most of us have tried the obvious fix: "write this email, but make it sound like me," with one example pasted in. Maddie's experience: "It sounded more human, maybe, but not really like me."

One prompt and one example can't capture a voice. Maddie first had to discover what her voice actually is, a process she compared to hearing yourself on a voicemail. "AI told me, you always open with a warm greeting. I'm like, ugh. Do I?" She did. Coming to terms with how you actually write, not how you think you write, is what makes the rest of the system work.

That's why her setup is six files, not one prompt: a voice identifier prompt, the voice guide it produces, a message analyzer, a "de-AI writer" that scrubs the output, a response generator, and a learning loop that keeps the guide from going stale by comparing what the AI drafted with what she actually sent.

How do I create an AI voice guide from my own writing?

The voice guide is the backbone of the system. As Maddie put it: "In order to de-AI something, you first have to get it in your voice. Otherwise, it's just going to sound like somebody else, still."

Her voice identifier prompt tells the AI it's a voice analyst and writing coach, with a precise goal: produce "a voice guide that someone else could use to write an email that I would read back and think, yes, that sounds just like me." It instructs the AI to look for patterns across samples, not one-off choices ("If it shows up every time, it's a trait of my voice"), analyzing opening and closing patterns, sentence structure and rhythm, punctuation habits, tone and warmth, the kind of humor you use (Maddie's is sarcastic and witty; no dad jokes), how you make asks, and whether you lead with the point or with context.

What you feed it matters as much as the prompt:

1. Use real writing you actually sent: emails, blog posts, or LinkedIn messages, never AI-assisted drafts.
2. Cover at least three contexts. Maddie included an email to a new manager, one to a potential client, and messages from different stages of a thread, since her voice warms up as a chain progresses.
3. Give it enough samples. "If you only give it one email, it's going to have to take everything from that email as the truth of how you write." Start with around five; twenty is even better.
4. Add a never-use list. Maddie keeps a list of words and habits she never wants, partly shaped by being warned throughout her career as a woman in tech about things like too many exclamation points. The system treats it as a kill list: "If it sees those, it removes it. That's it. Point blank."
5. Start with one channel. Her first version tried to handle emails, blog posts, and LinkedIn at once, and it wasn't great. Her Substack voice is genuinely different from her email voice. Get one working, then expand.

How do I get AI to answer what an email is really asking?

A perfectly voiced reply still fails if it answers the wrong question. That's the job of the message analyzer, which digs into the ask versus the need. Someone asking "when will the deck be ready?" often really wants to see it and give feedback. Someone asking for pricing may also want the timeline and whether there's wiggle room, or may just want pricing.

The analyzer reads the message closely (literal ask, subtext, emotion, urgency, what's missing), checks the thread history, researches the sender on the web if they're new, then synthesizes what they asked for, what they really need, and what would be most valuable. It has guardrails, too: no hedging and no psychoanalyzing. Read the message, not the person. As Maddie summarized: "A response, even if it's in your voice and it's not AI slop, doesn't go anywhere if you're not getting to the need and the root of the problem."

How do I strip the AI tells out of a draft?

Once a response is drafted in her voice, the de-AI writer runs a scrubbing loop, and the order of operations is the clever part. The first, mandatory step is a live web search for current AI writing tells, using three to five credible sources from the past six months: "AI moves too fast to look beyond six months ago."

Then it checks, in hierarchy order: Is the message still saying what I meant? Are there vague remarks that don't connect to anything real? Is my voice intact? Is the sentence structure a tidy formula or actually mine? Are the AI signature words gone? Only last comes punctuation, including em dashes, the simplest thing to fix.

Every flag gets checked against the voice guide. If the guide says Maddie genuinely uses a pattern, it stays. Real people do use groups of three, so the tricolon only gets cut when it feels automatic. Judgment calls get flagged back to her, with what was removed and why, so she can correct it or teach it. The result, from her live example (an email adjusting the start date at her new job), read exactly like her. "This is purely AI-generated," she told the group. For someone who "reads my emails 400 times before I send them, and then 400 more after I press send," getting the first draft to 90% changes everything.

How do I set this up in ChatGPT or Claude?

Maddie runs hers as a custom GPT she calls "Maddie's email assistant," with the six files uploaded as knowledge and instructions on how to use each one. Because it's connected to her Gmail, she can say "draft an email to Allison to reconfirm logistics for today's session" and it pulls the real thread. The same setup works as a Claude project or a Gemini Gem.

Practical tips from the Q&A:

• If the AI skips your instructions, mark steps as mandatory in the files, give thumbs-down feedback when it fails, and work inside the dedicated project or custom GPT rather than a general chat. Maddie's framing: treat it like an intern. Tell them to do five things, and "they're probably going to do four of them," until it figures it out.
• Keep instruction files short and specific. "AI is really good at high volume when it breaks it down into really detailed tasks, not so good at a really complex task all at once," Maddie said.
• Pair it with agents. An agent that flags emails needing replies can trigger this system directly.
• Keep a human in the loop. Maddie never lets the system auto-send: she reviews every draft and moves it out of the AI tool herself. For the right audience, she's comfortable saying a draft was AI-assisted and human-reviewed.

One wrinkle from the Q&A: members discussed emerging AI text watermarking, patterns encoded in word frequency that survive copy-paste. Maddie hadn't tested against it yet, but hopes a strong voice guide already pushes output toward your patterns, not the model's.

Key takeaways

• One "sound like me" prompt can't fix AI slop. You need a system: a voice guide from real writing, a message analyzer, a de-AI scrubbing pass, and a learning loop.
• Build the voice guide from at least three contexts of genuinely human-written samples (around five emails to start) and accept what it reveals about how you actually write.
• Keep a never-use list and treat it as a kill list.
• Answer the need, not just the ask. A well-voiced reply to the wrong question still fails.
• Scrub structure before punctuation. Em dashes are the easiest tell; formulaic patterns and hollow phrases are what give you away.
• Start with one channel, then expand. And always review before you send.

If you're still building your AI foundations, start with AI Skills for Non-Technical Professionals: Where to Start. And once your writing sounds like you, make sure it gets found: see Emily Richardson's session on how to show up in ChatGPT answers.

Join us for the next one

This walkthrough came from a live Willa event, where members saw Maddie's actual files on screen and got her voice prompt to adapt. There's already talk of a part two on the learning loop. Willa hosts expert-led sessions like this almost every week for women building careers with AI. See what's coming up on our events page, or become a member and learn alongside a community figuring this out together, in real time.

• AI writing
• voice guide
• AI slop
• prompt engineering
• custom GPT
• email

Join Willa, the community where women learn AI together.

All posts