Before it writes a word for you, Ad Machine has already read the ads that run for months. I took the number-one ad from the libraries of ten large brands, in health, beauty, pets, home and baby, and broke each one down sentence by sentence: what the line does, where the product shows up, how the ask is made. Those skeletons are in the box. My own two years of daily tests went in afterwards, losers included, and only where the big brands and my account agreed. Install it in Claude Code or Codex in an afternoon. It interviews you, pulls your buyers' exact words out of your reviews, writes from a proven structure, reviews its own draft, and remembers what you approve, so next month's batch can't repeat this month's mistake.
One-time payment, three editions, 30-day guarantee on all of them. Black Weeks prices until 30 November, no code needed.
Checkout isn't connected yet. Paste the Gumroad links into the CHECKOUT block at the bottom of this file, or set CHECKOUT_MODE to "stripe".
| Would the buyer stop, not just the curious | yes |
| One read: who, what's wrong, why this | clear |
| Objection answered in the opening | "I've heard this before" |
| Proof | [PROOF: sizer photo, real] |
Fictional brand, generated photos, no engagement numbers: a mock-up to show the method, not a measured result.
From the beta · four testers, free access, their own words
Quotes reproduced as written, with permission."I gave it a fair shot. My favorite part? It gives you something more useful than 'here's another version.' You can look at a specific line, understand what's weak, and work on it. That alone made this worth opening."
"Got to try this before launch. I thought I needed help writing headlines. Turns out I needed a better idea of what to say first. The research prompts helped me get there. Pretty happy with that."
"I was asked to test this out, hell... I wanted to! Having a script and video editor is nice. Having the shots, timing, and notes together is a lot more useful. I could actually picture how to make the thing. Really liked that part."
"Had early access to Operator. The worked edits were my favorite bit. Seeing the original, the change, and the reason helped things click. I found myself looking back at my own drafts and spotting stuff I'd skimmed straight past."
What the ads that run for months have in common
Not mine. Theirs. In November 2026 I pulled the number-one ad by impressions from the library of a large advertiser in each of five categories, two brands per category, and took every one apart: what each sentence does, where the product first appears, how the ask is made. Here is the sample.
| Category | What it is | Length | Days running | Rank in its library | Reach / library size |
|---|---|---|---|---|---|
| Baby | A mother's post, training pants | 4,800 characters | 335 | 1 of 291 | 840,000 people reached |
| Health | First-person letter, pain cream | 24,000 characters | 124 | 1 of 724 | library now 2,779 ads |
| Beauty | First-person letter, neck cream | 19,400 characters | 238 | 1 of 254 | library now 650 ads |
| Health | Founder answers "scam", gummies | 11,300 characters | 47 | 1 of 248 | 530,000 reached in 60 days |
| Hair | Voice-over video, scalp treatment | 98 seconds | 363 | 1 of 208 | — |
| Beauty | Creator demonstration, colour stick | 77 seconds | 179 | 1 of 1,479 | — |
| Home | Founder to camera, body pillow | 87 seconds | 303 | 1 of 180 | — |
| Beauty | Story video, face serum | 382 seconds | 61 | 1 of 278 | — |
| Pets | Owner's story, dog dental powder | 222 seconds | 182 | 1 of 220 | — |
| Health | Reaction video, joint supplement | 129 seconds | 189 | 1 of 368 | — |
Source: a public ad-intelligence library, read 7–8 November 2026. A long run and a number-one rank mean the brand keeps paying for the ad; they say nothing about its profit, and I don't guess at anyone's revenue.
A child's age in months. The exact treatments she's been offered. The vet's verdict. Not one of the ten opens on a riddle, and not one names its ingredient first. The hair video holds its molecule back until second 44.
After the inventory of what she already tried, with prices. After one line that says the failure wasn't hers. After a mechanism she can check against her own week. The pain-cream letter names its product three-quarters of the way down 24,000 characters.
"Link below." "Tag down below." The mother's post hides its link in an "EDIT: getting a lot of DMs". Then a P.S. that repeats the one belief that matters. Nobody ends on education.
Ad Machine writes from these skeletons. Your facts go into the slots: who speaks, what she tried, the number, the villain, the mechanism, the proof. Then one deliberate change, named. Zero changes is a copy that competes on their budget. Three changes is a guess. One is a test you can read.
The loop
That one is mine. Everybody clicked. Nobody who clicked wanted the product. I ran the loop for the best part of two years on two brands of my own, every day, and I can tell you where it goes: you blame the account structure, then the pixel, then the platform, then you buy a course, and the course makes you better at describing the problem without ending it.
What broke it wasn't a cleverer prompt. It was a file. Every dead ad got written up: what I tested, what I held fixed, what the numbers looked like, what I'd do differently. The next draft had to read that file before it was allowed to exist. Then I did the thing I should have done first and read what the brands running hundreds of ads were doing, because their test is the one my budget can't buy.
AI will do the boring work for you all day. The one thing it can't do is think for you. Let it decide and you'll never know why the winner won, and you'll be back in the loop the week it dies. Ad Machine is the file, with the writer attached and the ten skeletons already inside. It asks before it writes, it reviews what it wrote, and it remembers only what you approve.
Why does anyone pay $8,000 for a watch when the phone in their pocket tells the time? Because it's the one thing they own that reminds them what they can't buy more of. A losing ad costs you twice. The spend you can earn back. The month, you can't.
How it works
A short interview, once per brand. Real product facts, the buyer's situation, the offer, real proof, what has already run. Unknowns stay unknown.
Seven research jobs read the forums, reviews and competitor libraries for you. Then paste in your own reviews: it keeps exact phrases, where each came from, and the objection each reveals.
It picks the proven structure that fits your concept from the ten in the swipe file, says what it would take to use it honestly, and names the one flip.
Three different reasons to buy, each a chain: the situation, why it keeps happening, why this product answers that. Scored, with the proof each one needs.
For each concept: an image brief, the body, two openings, two headlines, a CTA. Short and long native versions when you ask. Video scripts in Copy + Video.
Ten checks before you see it. Then notes that quote the line, name the issue, give the reason and show the fix or the proof slot. Any of the twelve copy moves by number.
Edit, reject or correct a fact. It answers line by line, revises, and asks one question: this campaign, or this brand?
Only what you approve is saved, in your words, in that brand's folder. A rejection is never turned into a rule. Nothing crosses between brands. Operator Edition adds the operator pass on request.
Sample campaign · fictional brief
Brief: Stowly, a backpack built to fit the personal-item sizer used by budget airlines. Buyers: people who have already paid a gate fee once. Proof available: the bag's dimensions and a sizer photo. No reviews yet.
Buyer words (illustrative)
The angle it picked: the sizer at the gate. The villain is a bag sold as a "carry-on" that won't fit the personal-item box. The flip: the bag was designed from the inside of the sizer outwards. Skeleton: the founder's pillow video, in text.
Review notes
Review notes
Fictional brand, generated photos, no engagement numbers: a mock-up to show the method.
What's inside
Coreswipe-file-2026The ten top ads from the table above, broken down sentence by sentence into the job each line does, with a skeleton for each that you refill with your own facts. Long posts, one-minute videos, long story videos. The words stay theirs; the structures are yours.
Coreresearch-promptsSeven research jobs you paste into Claude or Codex: the customer corpus, who already knows what, the desires list with counts, the buyer-language file, who is scaling what, the structure file, the one-page brief. Setup for a browser extension so it reads full threads, not snippets.
Corehalbert-lessonsGary Halbert, Bond Halbert, John Carlton and Michel Fortin, rewritten in plain words and sorted by the job each lesson does in a feed ad: before writing, the argument, the opening, the close, testing and economics.
Corecopy-examplesTwelve editing moves from my own rulebook, each as a rule, a weak line, a strong line on a fictional product, and the market check. Qualify in line one. The four jobs of a hook. Absolve early. Question the familiar fix at a number she never checked. Reframe the price with a flip she already believes.
Corecopy-foundations · research-and-conceptsThe decisions behind an ad and the brief interview that gathers your facts, the buyer-language bank, three genuinely different buying arguments with the evidence gaps named.
Corewrite-and-review · feedback-and-memoryThe static package, long native copy and presells, the ten-check review that quotes the line, and the sequence that turns your edits into rules: accept, reject, correct; campaign or brand; saved only in your words.
Coretesting-basics · complianceOne variable per ad, your own break-even and loss cap, a table for reading the signals. Health, money, personal attributes, testimonials, AI people: a checklist, not legal advice.
Copy + Videovideo-ad-scriptsOne approved concept becomes a production header, timed scene table, two openings on one body, shot board, caption plan and edit checklist, for AI video tools or a crew.
Copy + Videoediting-passesBond Halbert's five editing passes, rebuilt for ads and scripts: mechanical splits, the read-aloud clarity hunt, the so-what and rhythm pass, format, final checks. Returns the edited text and the pass that caught each line.
Operatoroperator-review24 judgments, each with the call, the reason, the market check against a named top ad, what the masters said, the evidence class and the exception. The stop rule by break-even. The five shapes a losing ad's numbers take, and my Ads Manager column set.
Operatormarket-study · flip-firstThe ten ads beat by beat with runtime, rank and reach, and which judgment each one proves. Then, for each: what to keep, the one thing I'd change first, and how to run it as a test.
Operatorfield-notes · casebook · exercisesTwelve patterns from two years of tests, read against the market. Seven before/after cases, each with a market check. Two blind exercises with a sealed key. One complete example pass in which the buyer refuses two of my edits.
Every edition ships with blank brand templates (brief, buyer language, decisions, test log), a fictional demo brand you can run first, and a written guide with the exact first prompts for Claude Code and Codex.
Who built it
I'm Norbert. Several years buying media. Through 2025 and 2026 I worked on two supplement brands of my own every day as the creative strategist: writing the ads, launching them, reading the results, saving the ones worth saving, hunting for the bottleneck, writing better ones. Several hours a day of exactly the work this product helps you do.
For a long stretch of it I was in the loop. I couldn't have told you what I was testing, or why, or what the last test had taught me, or whether the next one came from research or from a hunch. Getting out meant writing those things down, every time, and reading them before the next draft.
My brands are small next to the businesses in the table above. One of those advertisers has 2,779 ads live in its library; another's top post has reached 840,000 people. I'd rather say that than pretend. So before any of my editorial judgments went into Ad Machine, I held each one against the ad ranked number one in those libraries, and against what Gary Halbert, Bond Halbert and John Carlton taught. Where the wider market agrees, the judgment says "market-checked". Where the evidence is my own small accounts, it says so. Where I'm acting on a hunch, it says "working hypothesis". You see the reasoning and decide whether it applies to your draft.
Core gives you the writer, the reviewer, the memory, the ten skeletons and the masters' lessons. My own preferences stay out of its defaults. The Operator Edition adds those judgments when you ask for them, with the reason behind each one, its market check, and the situations where it doesn't apply.
Operator Edition
Core writes from the market's structures and your facts. The Operator Edition adds the judgment, on request. The market's first: each of 24 judgments is held against the top ad of a brand with 180 to 2,779 ads live at once, ads that have run 47 to 363 days, one of them in front of 840,000 people. Then mine: two brands tested every day for two years, every dead ad written up. Then the masters': Gary Halbert, Bond Halbert and John Carlton, in plain words, sorted by the job each lesson does in a feed ad.
Ask for the operator pass and your draft comes back as a table: the line, my edit, the judgment, the reason, the market check, and whether it applies to your brief. You accept or reject each one. Nothing I think becomes a rule in your brand unless you approve it in your words.
Ten top ads, beat by beat, with runtime, rank and reach. Which judgment each one proves.
On each proven structure: what to keep, the one thing I'd change first, and how to run it as a test.
Every new file in all three skills, by email, for as long as the product exists. Only this edition.
Also inside: the field notes, a seven-case casebook with a market check on every case, two blind exercises with my answers sealed in a separate file, and one complete example pass in which the buyer refuses two of my edits. What it isn't: a promise of results, coaching, or a summary of anyone's book.
Specimen · one row of an operator passyour call
| Line | "Finally, a bag you can trust." |
|---|---|
| Operator's edit | "Fits when it's full. We photographed it in the box to prove it." |
| Judgment | J-07 People language, not a line every brand could sign |
| Why | "Trust" is what every bag says. The brand's claim is the condition, and the photo is the proof. |
| Market check | The top ads carry numbers the buyer thinks in ("$34 a bottle. Six bottles."), never adjectives. Pain-cream letter, 124 days, rank 1 of 724. |
| Evidence | market-checked · stated |
| Doesn't apply when | the brief has a real trust proof (a count, a certification) and nothing else |
| You | accept reject correct |
Fictional draft. Evidence classes: market-checked in the top ads · market-backed on my account · stated · working hypothesis.
The offer
CoreBlack Weeks
You save $50 until 30 November
The writer, the reviewer, the memory, and the ten skeletons.
Copy + VideoBlack Weeks
You save $50 until 30 November
Everything in Core, plus
For when a concept sells and you want to scale it: flip the motive, cut the variations, take the winner into video.
Operator EditionBlack Weeks · best value
You save $100 until 30 November
Everything in Copy + Video, plus
Bought a lower edition and want more? Reply to your receipt and pay the difference between what you paid and the current price.
30-day guarantee, all three editions. Give it one real brand and let it write one batch. If what comes back isn't better than what you'd have written without it, reply to your receipt with that batch within 30 days and I refund the full price. Why the batch? A zip can't be returned, so the guarantee is for people who ran it, not people who downloaded it. I don't grade your batch and I don't argue with your verdict. It's a tool to do better work and learn faster, not a promise of a winning ad; nobody honest can promise you that.
Unsure which edition? Start with Core and upgrade for the difference when you've seen it work.
Payments and VAT are handled by Gumroad. You get the download link by email within a few minutes.
Secure payment by card, Apple Pay, Google Pay or PayPal, powered by Stripe. Your download link lands in your inbox within minutes. 30-day refund policy.
Customer reviews
Every review here comes from a buyer and is posted automatically, exactly as written. I don't edit reviews and I don't hide bad ones.
More from the beta
"I got free access to test this. The part I kept coming back to was the explanation behind the edits. It helped me understand why an opening was weak and what information was missing. That's the stuff I usually spend ages trying to work out myself."
"I got a free copy to try. My favourite part was being able to push back on the feedback. I could explain what I disagreed with, get another version, and save the preference for the next batch. That back-and-forth made it useful."
"I had free beta access. There's a lot of thought in the questions it asks before writing anything. It made me slow down and decide what the customer needed to believe, what proof I had, and what I was actually trying to test."
Beta testers were given the product free in exchange for honest feedback. Quotes are reproduced as written, with permission.
Questions
Claude Code, or OpenAI Codex (the CLI or IDE extension, running on GPT). It also installs into claude.ai and the Claude desktop app as uploaded skills. Installing is unzipping one folder. The guide in the zip gives the exact first prompts. It isn't a Custom GPT and we don't promise it works in the ChatGPT web chat.
November prices: Core $97 instead of $147, Copy + Video $147 instead of $197, Operator Edition $197 instead of $297. No code; the lower price is already on the button. At midnight on 30 November the page and the store go back to list prices, and the countdown above the cards is the real one.
Core if you write your own ads and want the skeletons, the research jobs and a reviewer that learns your brand. Copy + Video if you make video or want Bond's editing passes on your scripts. Operator if you want a second opinion with the market checks attached, and lifetime updates. You can upgrade later for the difference.
Because their test is bigger than mine. A brand with 2,779 ads live keeps paying for its number-one ad because it works; my account can't buy that evidence. My judgments are in the Operator Edition, each marked by how sure I am and checked against those ads. Where I only have my own account to go on, it says so.
Each judgment says how sure I am: market-checked (the top ads of large brands do the same thing; the market study names the ad), market-backed (my account data pointed the same way, on small samples), stated (I made the call and apply it), or working hypothesis. Treat them as a disciplined starting point, not a law.
All three editions, for 30 days from purchase. Run it on one real brand. If the batch isn't making your ads and your judgment better, reply to your receipt with that batch for a full refund. Haven't had time to run it? Ask before day 30 and you get 14 more days. It's a tool to do better work and learn faster, not a promise of a winning ad. The full rules are on the refund policy page.
Operator Edition buyers get every new version of every file, by email from the store, for as long as the product exists. Core and Copy + Video are the version you buy, complete. Upgrading for the difference moves you up at any time.
One folder per brand. The skill loads only the brand you name and never carries facts, phrases, examples or rules between brands.
No. When a line needs proof you haven't given it, it leaves a visible slot such as [PROOF: number of customers] and lists it under "Before you publish". Your name goes on the ad, so the facts have to be yours.
Yes. It reads a compliance checklist before writing in those categories: personal-attribute wording, health and income claims, testimonials, AI-generated people. It's a checklist, not legal advice. Several of the top ads in the swipe make health claims that would fail it; the swipe is about structure, not claims.
A free skill writes from your product description and starts from zero every chat. Ad Machine starts from ten proven structures and your buyers' words, reviews its own draft, and keeps a per-brand memory of what you approved. Week ten writes closer to what wins on your account than week one.
It writes in your market's language from that market's buyer words, rather than translating English ads. It has been tested in English. Polish and other languages work, but check the first batch closely.
No. It writes the ads, explains what it changed and why, and learns from your edits. The lessons inside are applied to your draft, not taught in videos.
You do. The copy it writes for your business is yours to use.
Free skill
The Buyer Words Test is one working skill from Ad Machine. Paste your ad and 20–100 real reviews or comments. You get the phrases your buyers actually use, a verdict on your opening, the opening rewritten in their words, and three new hooks.
Free download. Just your email.
YOUR OPENING: WEAK Why: line one names the product, not the moment. A curious stranger clicks as fast as a buyer. REWRITTEN OPENING "It fit on the way out." 3 HOOKS FROM YOUR BUYERS' WORDS 1. I paid more for my bag than my seat. 2. …