How to Reverse Engineer a Winning Ad Script With ChatGPT or Claude

Short answer: Find a video that is already working in a nearby category, transcribe it, and break it into its structure: the hook pattern, the credibility line, the order of the reasons, the bonus and the close. Then give an AI model that structure plus the facts from your own product page and ask it to rebuild the script for your product. Add a style pass so the new script keeps the rhythm of the original. You borrow the skeleton and the pacing, never the words or the claims.

The brands in the examples are made up.

What does reverse engineering an ad mean?

Reverse engineering an ad means taking a script that already holds attention and separating what makes it work (structure, order, rhythm, length of each beat) from what it says (the product, the facts, the wording). You keep the first and replace the second.

It is the same thing good copywriters have always done with swipe files. AI just makes the rebuild take minutes instead of an afternoon. It pairs naturally with ad library research: the library tells you what is running long, this process turns it into a draft for your product.

Why does this work better than writing from scratch?

A script that is already working has been tested by the market. Its hook stops people, its order keeps them watching and its close gets the click. When you write from a blank page you are guessing at all three.

The structure is usually what carries the result, not the specific words. A "here are four reasons I'm obsessed with this" list, said by someone with a clear credibility line, works across finance, food, home and beauty because the pattern itself is easy to follow. That is processing fluency doing its job.

What is the step by step process?

  1. Find the source. Look for a short video in a nearby category that is clearly working: long running in the Ad Library, or high views and saves on a creator's organic account. It does not need to be your category. Often it is better if it isn't, so you are not copying a competitor.
  2. Get the transcript. Screen record the video and run it through a transcription tool (we use Descript), or paste captions if they are available. Clean up the obvious transcription errors.
  3. Map the structure. Write down the beats in order, with a rough length for each: hook line, credibility line, reason one, reason two, and so on, bonus, close. Note the hook pattern from our hook templates.
  4. Collect your facts. Copy the real claims, prices, guarantees and FAQ answers from your own product page. This is the only place the facts in your new script are allowed to come from.
  5. Run the rebuild prompt. Give the model the original transcript, your structure notes and your product facts, and ask it to rebuild the script for your product.
  6. Run the style pass. Ask the model to analyze the rhythm of the original and match it.
  7. Edit by hand. Cut anything that is not true of your product, check every claim against your page, and read it out loud. Then hand it to a creator with a creator brief.

What prompts do we use?

The rebuild prompt:

Reverse engineer the script below for [your product]. Keep the structure, the number of reasons, the order and the length of each beat. Replace every fact with facts from the product details I give you. Do not add any claim that is not in the product details. Script: [paste transcript]. Product details: [paste product page copy and FAQ].

The style prompt:

You are a stylistic forensic expert. Analyze the original script for sentence rhythm, punctuation density and average word complexity. Then rewrite the new script so it matches all three.

The style pass matters more than it sounds. A rebuilt script often comes back longer and stiffer than the original. Matching sentence length and word complexity brings back the spoken, easy pace that made the original watchable.

What does it look like on a real example?

Say the source is a creator video in personal finance that runs: "Here are four reasons I'm officially obsessed with [product]. I'm a former [credibility] and now I make free content so you don't have to pay for it. Number one... Number two... Number three... Number four... And if this was helpful, like and follow."

Mapped, that is: list hook, credibility line, four short reasons with one fact each, a bonus, a soft close.

Rebuilt for Copperbean, using only facts from its own product page:

Here are four reasons I'm officially obsessed with Copperbean. I make coffee every single morning, so I've tried basically everything. Number one: it's brewed slow and filtered twice, so it's smooth. No bitter aftertaste. Number two: one bottle makes a week of coffee at home. No more drive thru line. Number three: it's a concentrate, so I make it as strong or as light as I want. Number four: it shows up on my doorstep before I run out. Bonus: it works hot or iced. If you've been paying café prices every morning, try a bottle first.

Same skeleton, same pace, entirely different words and facts.

How do you keep this ethical and safe?

  1. Borrow structure, not sentences. If a line from the original survives the rebuild word for word, rewrite it.
  2. Never borrow claims. Every fact must come from your own product page or your own data. AI will happily invent a statistic if the original had one in that spot. Delete it.
  3. Stay out of your direct competitor's ads. Pull structures from neighboring categories so your ad doesn't look like a copy.
  4. Check regulated categories twice. Finance, health and supplements have rules about what you can claim. The original creator's claims may not be ones you can make.
  5. Keep the credibility line true. If your creator is not a former financial advisor, they don't say they are. Swap in a credibility line that is real ("I've been making coffee every morning for years").

How do you know if it worked?

Test the rebuilt script like any new concept. Judge it on stop rate and hold rate first, since those tell you whether the borrowed structure is doing its job, then on outbound click rate and purchases. Our article on reading a video ad in four stages walks through each. If the hook stops people but the hold rate drops at reason two, the structure is fine and the reasons are weak. Swap in stronger facts, not a new structure.

Key takeaways

  • Start from a structure that already works, not a blank page.
  • Transcribe the source, map its beats, and collect your own product facts before you prompt.
  • Use a rebuild prompt and a style prompt so the new script keeps the original's pace.
  • Borrow the skeleton and the rhythm, never the words or the claims.
  • Pull structures from nearby categories, not your direct competitors.

FAQ

Is reverse engineering an ad the same as copying it? No. Copying keeps the words. Reverse engineering keeps only the structure and pacing and rebuilds everything else around your own product and facts.

Which AI tool is best for this? ChatGPT and Claude both handle it well. The prompts matter more than the tool. Always include your real product details so the model has true facts to work with.

Where do I find scripts worth reverse engineering? Long running ads in the Meta Ad Library and high performing creator videos in nearby categories. Our guide to Meta Ad Library research covers how to spot them.

How many versions should I make from one source? Start with two or three: the straight rebuild, one with a different hook from our hook swipe file, and one with a different credibility line. Then let the test decide.

Want a second set of eyes on your scripts?

We'll look at what's running in your account and where your creative is leaking attention, for free. You get a full account audit plus net new reach and cumulative reach reports. Request it on our contact page.

Check what you learned

Tap an answer to see if you got it right.

1. Where should you pull structures from?

Nearby categories give you proven structure without copying a rival.

2. What do you borrow from the source ad?

Never borrow the words or the claims.

3. What do you give the AI to rebuild the script?

Map the beats first, then add your real facts.

Ready for the full test? Take the Creative exam and get certified. Want this run on your account? Get my free audit.

About the author

Jake Mehani, founder of Creative Analytics

Jake has spent 15 years running Meta and paid social ads for consumer brands. Meta published his RBX Active Advantage+ case study, he taught the DMAnc Meta and TikTok ads masterclass, and he spoke at Motion's 2024 Creative Strategy Summit. His senior team runs media buying, creative and measurement like an in house team.

Jake on LinkedIn  |  Case studies  |  Free Meta audit

Previous
Previous

Flat vs Compounding Brands: RFM Segments and the Four Numbers for Your Quarterly Plan

Next
Next

Remarketing Is a Reminder, Not a Sale: Reach vs Conversion Objective on Meta