I worked in search marketing many, many years ago. Later I spent time in Indeed, who were incredibly good at search marketing. They took it to a whole new level. I had previously worked at brands where search was a tactic delegated to more junior marketers. In Indeed, it was central to their business and they treated it like a science.

Since AI became a thing, I’ve got very interested in search again, as the customer journey started to change dramatically.

What happens when our buyers are asking ChatGPT, Claude and Gemini before they ever reach our websites? I’ve spent a lot of time getting under the skin of this, reading, listening to experts, experimenting myself. I’ve written about what I’ve learned in my essays and book.

So if this is something you’re spending time on - there a free HubSpot masterclass on it this Wednesday (19th August, 2pm BST / 3pm CEST) from folks more knowledgeable than myself.

Beeri Amiel is presenting. He co-founded XFunnel, the technology HubSpot bought and turned into their AEO tool for mid-market and enterprise teams. So you're hearing it from someone who built the thing, rather than someone who read about it.

They're going to run their AEO Grader live in the session. You put your brand in and see how ChatGPT, Perplexity and Gemini are describing you, and where you sit next to competitors.

It’s free but you need to register.

Enjoy.

Paul

If you’re interested, pop over to my website. Or share with folks you think would like it.

Any talk of AI replacing marketers is nonsense. But marketing roles will change a lot, and for the better. And I’d love as many marketers to embrace it and become fluent.

Time is the biggest blocker, so I’m hoping this course will help jumpstart marketers and show what is possible. I’ll add a bunch of things to try, copy and keep. I’ll share lots of links and point folks towards experts that I think are worth keeping an eye on.

The course opens on Monday 21 Sept. You can do it in your own time. I’ll add each module every two weeks and then leave them there for anybody that wants to binge on them later.

runwithfoxes.com

Segmentation is one of the fundamentals of marketing. Almost every piece of strategic work starts by deciding who we’re for, and that means deciding that some groups of buyers differ from others enough to pause and think about how to target each. Few argue with that.

Yet flabby thinking could lead to very serious mistakes. I’ve been in meetings where business consultants have produced very polished presentations showing why we should ignore some segments, and double down on others. But segment differences need to be significant to be useful. And you need to be able to target them with different product variants, media, messaging. Otherwise it is just theory.

Because segmentation is so engrained into how we learn about marketing, we can fall into the trap of looking for differences that don’t exist. If you start out already committed to finding differences, I guarantee you’ll find them.

If we start assuming we have segments, we build our surveys to separate people, we build the charts to show the separation, and by the time the final deck exists we’ve already committed to the answer. This is hedgehog thinking.

A fictional example to demonstrate

Look at the chart. At first glance, we have five customer segments scored on six typical brand statements. We can see some serious differences here. Five distinct segments. Five different shapes, five different audiences. Some people think the brand has wider range of products that others. Some trust it more. Some find it less easy to buy.

Great, so this gives the starting point on how to succeed with each segment if we want. Or at the very least, design our plans to target these different groups.

Except maybe we don’t. Here is the same data, with a proper axis. Axis at 0-100. The spread was three or four points all along. So we don’t have five segments based on this data. We have one.

We draw differences because we went looking for them. Crop the axis, index the scores to 100, show one segment per page so nobody overlays them, and four points of noise becomes a finding. Every chart is accurate. The framing is the trick.

This fix is almost embarrassingly simple. Take a segment, and for every statement, measure the gap to the average shopper and drop the sign, so a plus five and a minus five both just count as five. Average those gaps across all the statements. Subtract from 100. That’s the segment’s similarity. A group sitting three points off the average on a typical statement comes out 97% similar.

The averaging is the important bit. One dramatic statement now has to share the average with every statement where the group is identical to everyone else. The cropped-axis chart let one spike speak for the segment. This won’t.

Andrew Ehrenberg found that brand buyers differ by under two points on average across dozens of categories. So the rule of thumb: clear 89% similarity and you are looking at one audience, not several.

Here’s the part that’s new, and the reason I’m writing this.

Segmentation matters, but pseudo-science segmentation is dangerous. And the maths that tells them apart has been around for decades. I’ve known about since I first started studying the work of Ehrenberg-Bass 15 years ago.

But knowing it was never the bottleneck. Doing it was. Running the sums properly, every time, used to mean a specialist, which meant time and money, which meant the check got skipped unless someone was pedantic enough to insist. I am. But pedantry doesn’t scale.

But if you know the fundamentals, you can use AI to do the computing. Some of the segmentation checks I built are below to see if the data shows if we really have differences in our segments.

And correlations checks to see how closely segments agree (not differ).

And how far away from average is each segment.

And how much of the differences are due to the brand size affect.

And how much of the differences is just due to purchase frequency - those they buy more generally say nicer things about the brand.

My point is this. Don’t settle for having deep knowledge. Combine your knowledge with AI to build things that create adjacent value.

Subscribe now

So the good news is that the way to win in SEO is something marketers know a lot about - fame strategies. The bad news is that fame strategies are hard.

Every few weeks someone declares SEO dead and sells you its replacement.

The current narrative goes like this. Search has split in two. There’s the old game, backlinks and Google rankings. And a new one running alongside it: get your brand mentioned across the web, even without a link, and AI models start citing you in their answers.

The number doing the rounds comes from Ahrefs, who looked at 75,000 brands suggest that brand mentions correlate with AI visibility about three times more strongly than backlinks, 0.66 against 0.22.

The top things are brand mentions, brand anchors (your name used as the words in a link), and brand search volume. Brand search volume is the number of people who type your name into a search bar - a count of the people who already know you exist and go looking for you. They measure - who gets talked about, who gets named and who gets looked for.

That isn’t about SEO optimisation. Nope. It’s about fame.

An AI answering “what’s a good running shoe” isn’t grading your website. It’s reaching for what the internet repeatedly says about running shoes. Repetition reads as credibility coming up across reviews, forums and videos, so the model treats you as worth naming.

It’s the same reason you trust a restaurant three separate people mentioned without being asked. Nobody showed you a backlink audit. You just kept hearing the name. So the thing that gets you cited is the thing that was always hard - getting talked about.

Asking brands, websites and owners of tools to mention your brand (even with no link) probably works at the margins, and there’s no harm in it. And the page-work still matters. If a model can’t read your site, or your claims are vague, or there’s no clean sentence to lift, you lose even when you’re known. Making your pages easy for a machine to read is worth doing. It’s probably table-stakes though.

When Ahrefs extended the study to ChatGPT and Google’s AI, the important signal wasn’t connected to website optimisation. It was youtube mentions. People talking about you on camera because you were worth the airtime.

You have to earn that the hard way - by being interesting enough that someone wants to talk about you for free. Purple Cow stuff (for Seth Godin fans).

None of this is new. Get known. Stay easy to find. Be worth mentioning. Peter Field and Paul Feldwick have been saying this for years. AI didn’t create a new discipline. It rewards the brands people already talk about, and it dragged brand fame into a spreadsheet the SEO folks are now forced to read.

Lovely to see marketing work coming together instead of being sliced and diced into a hundred different teams.

Subscribe now

In fairness, this was changing before the robots settled in:

• SparkToro and Datos put US zero-click Google searches at 58% back in 2024.

• Pew found that people clicked a link 8% of the time when an AI summary showed, against 15% when it didn’t.

• Ahrefs has the number-one result losing 58% of its clicks when an AI Overview sits above it.

• BCG and Moloco’s research suggests 33% of US adults now discover brands through personal AI agents, and 47% already use AI tools to research purchases.

Sources: SparkToro & Datos, 2024 · Pew Research, 2025 · Ahrefs, 2026 · BCG × Moloco, 2026

So the category got a name. GEO (Generative Engine Optimisation) and a queue of people who updated their LinkedIn bio to match. Some of it is genuinely new. A good deal of it is the SEO we should have been doing anyway. So here the list of 18 things worth knowing about GEO:

1. Ranking still feeds the answer

Google’s own guidance says SEO best practices “remain relevant for AI features in Google Search,” and that “there are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary.” AI Overviews are assembled from the normal search index. If you do not rank, you are not in the raw material. This part has not changed.

Source: Google Search Central, AI features guidance

2. The AI can only quote pages Google already knows about

Before a page can turn up in an AI answer, Google has to have it filed in its index, the catalogue of pages it has crawled. If your page isn’t in there, the AI can’t see it, so it can’t quote it. Google’s free Search Console tells you which of your pages are filed and which are missing. Fix that first before we invest in other AI tactics. It’s plain old SEO.

3. Check you are not blocking the robots

Cyrus Shepard went through 54 separate AI-citation studies, patents and experiments and ranked the factors. The one at the top was dull: can the page be crawled at all. Sites often block the AI crawlers without realising.

Source: Cyrus Shepard, Zyppy Signal

4. A lot of the “special GEO files” do nothing

Google, in writing: “You don’t need to create new machine readable files, AI text files, or markup,” and “there’s also no special schema.org structured data that you need to add.” So the llms.txt file someone is trying to sell you does nothing for Google. Keep schema for your rich results if you want it, but not as a magic word for the machines.

Source: Google Search Central, AI features guidance

5. Human writing still wins the ranking that feeds the answer

Semrush looked at 20,000 keywords and found human-written content in the top spot 80% of the time, against 9% for content it flagged as purely AI. Their own caveat is: AI detectors are unreliable, and Semrush sells SEO tools. Take it as a lean, not a law.

Source: Semrush, AI content ranking study, 2026

6. Move from the page to the sentence

Here is a proper change from SEO. A search engine ranked pages. A model lifts a sentence. It does not cite your site, it quotes one clean line and moves on. So the question is not “is this page authoritative” but more like “can one sentence survive being pulled out and dropped into someone else’s answer.”

7. Be worth quoting

The original academic work on this, a Princeton paper from 2024, tested what nudges you into AI answers. Adding statistics, quotations and cited sources all moved the needle, by up to 40% in their words. Keyword stuffing made things worse, not better. So say specific, sourced, quotable things.

Source: Aggarwal et al., “GEO: Generative Engine Optimization”, KDD 2024

8. Answer early, and stop padding

Dan Petrovic’s work on how Gemini grounds its answers found about a third of a page gets considered at all, and a tight 800-word page got over half its content used against roughly 13% for a 4,000-word one. Put the answer near the top and let the page be short.

Source: Dan Petrovic, DEJAN

9. Write headings that match the question

AirOps looked at 16,851 queries and found pages whose headings closely matched the question were cited 41% of the time, against 29% for weak matches. Phrase the heading the way a person would ask it out loud, not the way a content calendar would file it.

Source: AirOps, The Fan-Out Effect

10. The bulk of what AI says about you is not on your site

AirOps again: across the brand mentions they studied, 85% came from third-party sources, not the brand’s own pages. Reddit, reviews, comparison posts, YouTube. You can write the best page in the world and the model will still reach for what other people said about you. The public record is the battleground, and barely any of it is yours. None of that is SEO as we remember it.

Source: AirOps, off-site signals in AI search

11. Off-site is the actual work, especially in considered categories

Which means the job spills well past your own site. For a bank or a B2B tool, the answer a model gives is being assembled from comparison sites, a forum thread from 2023, a handful of reviews. If that record is light, stale or confused, that is your exposure, and no amount of on-page polish fixes it.

12. The engines do not agree with each other

Back in early 2025, Seer found 87% of ChatGPT’s search citations matched Bing’s top results, and only 56% matched Google’s (a small sample, and the products have moved on, so treat it as directional).

Source: Seer Interactive

13. Citation is a probability, not a position

A rank holds still. An AI citation does not. Ask the same question twice and the sources can change. So the measure is not “are we number one?,” it is “how often do we show up across a basket of the questions our buyers ask, tracked over time?” Anyone promising a fixed spot is selling certainty that does not exist. [Note: More of an opinion than an empirical study].

14. Getting cited is worth real traffic, but keep it in proportion

Seer’s 2026 data: when you are cited in an AI Overview you get 120% more organic clicks per impression than when you are not. The footnote, which Seer make themselves, is that a result with no AI Overview at all still beats both. So a citation is the good outcome in a world with fewer clicks going, not a windfall that hands the clicks back.

Source: Seer Interactive, 2026 AI Overview update

15. Do not build a page per question

The tempting move is to spin up a page for every variation a model might ask. Google names that as scaled content abuse, and it is an own-goal regardless. Put the answers into the pages you already have.

Source: Google Search, spam policies

16. Make the boring pages do the work

The help centre, the pricing edge cases, the compatibility notes, the FAQ you buried three clicks deep. This is now very good raw material. A plain sentence in a support doc gets quoted where a glossy homepage gets skipped. Your About and credibility pages are the cheapest win, because they are how a model works out who you even are.

17. Pages that finish a job beat pages that only explain

Cyrus Shepard went through hundreds of winning and losing sites and found the survivors do more than inform, they get something done. His strongest performers are the pages where the reader actually does the thing: a transaction, a tool or calculator that turns your input into an answer, a forum where real people settle real questions. A model cannot hold the community, or take the booking, so if you have these, you’re in a better position. Specific, first-hand, useful content gets cited. The commodity stuff anyone could generate does not.

Source: Cyrus Shepard, Zyppy Signal

18. Where this is heading: loyalty, then agents

Two things on the horizon, both early. Google has started feeding the preferred sources a reader picks into AI Overviews, which turns earned loyalty into an input for AI visibility rather than a nice-to-have. And agentic search, where an assistant does the buying for you, is on Google’s own roadmap. It is coming but could be a while. Build for today’s reader and you are building for both.

Source: Google, Preferred Sources in AI Overviews

Subscribe now

None of this is a dark art. It is being findable, being worth quoting, and being talked about in the places the machine already trusts.

Done. Live.

Subscribe now

The interesting part of this, is what’s missing.

There’s no CMS. No admin screen, no plugins, no Squarespace workspace where the big creative decision is which template comes closest to the thing in your head. I killed all of that.

The website is code. Claude runs on my laptop and can read and write that code. A folder of files that lives on GitHub, which keeps a copy of every version that has ever existed. It is hosted on Vercel but Claude talks to it, not me.

I press a button, I talk, it types, designs and deploys. That’s the whole machine. None of it is exotic or expensive. Speed without quality is just a faster way to embarrass yourself, so the quality lives somewhere specific: skills. A skill is a brief written down once, properly.

My brand skill holds the colours, the typefaces, the spacing rules, what a hero section is allowed to do, how the chatbot speaks, what good UX looks like on a phone. When I ask for a new page, Claude Code reads that skill before it types a word. So I’m never asking an AI to guess what good looks like. The guessing happened months ago, with me in the room. On the day, the sentence is just an instruction.

And once the machine is in place, the relationship with your own website changes. It stops being the thing you’re afraid to touch and starts becoming an active part of your marketing.

If you want a landing page, say you want a landing page. If a new client needs a private section behind a password, ask for that. If the home page hero feels static, make it a rotating hero. Gate a piece of content. Ungate it the following week when you change your mind.

If you want ten chatbots, one for each section of the site, you can have ten chatbots. I’m not saying that’s a good idea. I’m saying it’s an afternoon. The small stuff barely registers. More space between two sections. A new line in the chatbot. Show your new product. Or your first book. Bestseller by the way…

A case study added the day it happened. Each one is a few sentences. None of them is a project.

There’s a safety net under all of this too. Because the website is code on GitHub, every change is a saved version. If I try something and it’s wrong, I roll it back to how it was an hour ago. Experiments become cheap. And cheap experiments are how websites get better, because the usual reason a company site stays mediocre is that touching it feels dangerous.

There is a cost. Not money. I can build a website in a few days.

The costs is giving up old website. A sunk cost. The instinct is to keep the CMS and bolt the AI onto the side, given you just redesigned the thing. But that won’t work. Bite the bullet and kill the old technology first. Before your competitors do.

Dump the website. Start again.

Subscribe now

Start with 95:5. At any given moment, only a small share of the people who could buy from you are in a position to. Roughly 5%. The other 95% have no immediate need today.

Why? They have an existing contract, or a supplier that’s fine for now, or their budget is already spent, or they are too busy with other priorities.

Subscribe now

The 95:5 principle comes from the wonderful and smart folks in The Ehrenberg-Bass Institute, this one from John Dawes. He had a great book for anybody interested. The crux of this principle is that if companies in a category change supplier about every five years, then in any given quarter about one in twenty are in the market.

Now the Day One list.

Bain and Google asked about twelve hundred business buyers how they choose. Before they started looking, 86% already had a short list of suppliers in their head, often just three names. That’s the Day One list. By the time the buying was done, 92% had bought from that original list. People do more research, most add a name or two, and then they buy from the names they walked in with anyway.

So put the two together.

1. Hardly anyone is in the market at once, and

2. The few who are mostly buy from a shortlist before they started properly searching.

So the decision is half-made before a buyer ever raises a hand. You’re either already in their head, or you’re not in the running. You can have the better product and the keener price and still lose, because the choice narrowed before you knew it had started.

Don’t ignore the 5%. Sweat that marketing as much as you can. More on that here.

But don’t focus your entire marketing efforts on it. It’s a small group, a different small group each quarter, and everyone else is fishing the same pond. The bigger prize is the 95% who aren’t shopping yet, because they’re the ones writing next year’s Day One list right now, from memory, and you want to be on it.

So our marketing has two jobs, not one. The first is to catch the 5% who are ready. Make it easy to find you, easy to act, easy to talk to a human or buy.

The second is to get onto the Day One list of the 95% who aren’t ready yet. That’s brand building. It’s the work that puts you in the buyer’s head a year before they need you, so you’re on the shortlist they bring to the table. You measure the first job this week. You measure the second over years, and judging it on this quarter’s leads is a poor decision.

We don’t know which 5% will be in the market next quarter. So chasing only the people showing buying signals leaves us invisible to the buyer who hasn’t started looking yet, which is the exact moment the list gets written. Generally, a good approach is to reach the category broadly and consistently, lightly and often, rather than heavy bursts followed by silence. You’re planting memory in people who won’t need you for another two years.

So what does this ask of you. A few things.

• Know your cycle. Work out roughly how often your customers buy, because that tells you how small the ready group is and how big the future pool is. More on the sum in a second.

• Split your money across both jobs. Some to catch the ready, more to build memory in the rest. The exact balance depends on your business, but if every euro is going to this quarter’s buyers, you’re starving the long game that feeds it.

• Be easy to remember. A brand that looks and sounds the same everywhere, with assets people know without reading the logo. For me that’s the foxes. It saves me explaining who I am every time.

• Be useful to the people who aren’t ready. Give them something now that asks for nothing back. I leave the first part of my book ungated. No form, no work email required. They take it, they remember where it came from.

• Make it easy for the ready ones. When someone is in the five percent, every extra step loses a few of them. Clear contact, a fast response.

You can do a rough version of your own 95:5 on the back of an envelope. The question is simple: in the window you care about, what share of your customers are at a buying moment?

Start with how long they go between purchases. Say they replace their supplier every five years, so sixty months. Now picture six thousand of them, with their renewal dates spread evenly across those sixty months. In any one month, a sixtieth of them come up, so a hundred companies. Look at a single quarter, three months, and that’s three hundred in the market. Three hundred out of six thousand is 5%. Look at a whole year and it’s twelve hundred, which is 20%.

The shortcut is the window divided by the cycle. Three months over sixty is 5%. Twelve over sixty is 20%. Period over cycle.

It’s rough, because it assumes purchases are spread evenly and that being due to buy is the same as being in the market. Neither is quite true. But you’re after the shape of your market here, not a figure to two decimal places.

The envelope sum is rough on purpose, and the honest version is fiddlier. That’s where I got AI to help.

Using Claude Code, I built a proper model for this. First it splits the market up. Big customers and small ones buy on different clocks, so a single average hides the truth. The model treats the groups separately and weights them, and the real in-market figure may come out higher than a flat 5%.

Then it works the funnel. Being in the market isn’t the same as buying from you. Some people never hear of you, some won’t consider you, some you never get a sales call into. The model takes the demand and strips it down through those leaks to a figure you could realistically win. That gap, between what’s in the market and what you’d get, is the useful bit. It also shows what the brand building is worth over the next couple of years, not just this quarter.

Then there’s the Monte Carlo part. When you forecast, you normally feed in one number for each thing. One buying cycle, one win rate, one deal size. The trouble is you don’t know those numbers for certain. The cycle might be four years or seven. The win rate might be 15% or 25%.

So the model takes a range for each one. Then it runs the whole calculation thousands of times, and each run it picks a slightly different number from inside each range, like rolling dice. You end up with thousands of answers instead of one, and you look at the spread. Rather than “you’ll win three hundred and sixty grand”, you get “probably between sixty grand and nine hundred grand, most likely around two hundred and sixty”. It’s named after the casino because that’s the trick, roll the dice enough times to get better information.

That spread is useful. A single forecast number is a guess. The range tells you how much faith to put in it, which is the thing the boardroom needs and rarely gets. The model is straight about what it’s sure of too. Where I’ve given it a real figure it uses it, and where I haven’t it fills the gap with a sensible default and marks it as a guess, so nobody mistakes an assumption for a fact.

This isn’t AI as a magic trick. It’s the boring, repeatable arithmetic handed to something that doesn’t mind doing it for the thousandth time, while the judgement, what the numbers mean and what to do about them, stays with me or the client.

None of this is a reason to spend less. It’s a reason to focus on the 95% and the 5%. Catch the few who are ready, and get your place on the list before they need you.

Unfortunately most of your target customers are not ready today. The whole game is being the one they think of when they are.

Subscribe now

Subscribe now

This ad is from an article I’m writing. When I have my original or hero asset, I can then point Claude Code at it and ask it to re-size it into the formats I need. Claude knows what sizes I want and in seconds, it spits them all out, in static, and Gifs and mpegs. The files done and ready to put live.

Claude is not drawing the ads. It opens each ad as a tiny web page and takes a picture of it frame by frame. Then it uses two free tools, ffmpeg and gifsicle, and turns those frames into a video and a looping GIF. That’s it. No image generation. No internet. Not even Python. These are just tools.

The real work is before that. To set up a system like this means taking everything you know about how you want this type of ad to behave, pulling it apart, and writing it down as rules.

For example, you need to think about what Claude should put into vertical ads. Or really small ads. If not, it will try to just take everything from the big hero ad, make everything smaller and shove it into a different size or shape. That doesn’t end well.

The brand guidelines must work every time. This is the important part, and is why I’ve not gone for generic solutions. I want the ads to be exactly right every time, without me getting involved.

So for the example above, stuff like the data in the chart should be blue. The hero number, and the “run” in the logo, is always orange. They never swap. Headlines never run to the edges. They sit at about 72% of the width, because type crammed edge to edge looks cheap. And every size ends the same way, closing on runwithfoxes.com with the logo fading as the address takes over.

None of that is the ad. It’s the knowledge of how the ad works, taken apart and made explicit. Once it’s written down, the machine just follows it. Every one of those decisions lives once, as a rule, in one place. The ad I feed in carries only the data, the numbers and the words. It carries nothing about the look. So the same ad in, gives the same pixels out, every time, on every size.

Once built, I couldn’t make it off- brand even if I tried. I’ve tried plenty.

The other really important bit for me is this. I’ve found that the more narrow the use case the better. So this is not about making one tool or system that works for all brands. Instead, it is about making a system for just one brand, and building that brand exact details into it.

So, build slowly and carefully first. Then you get speed.

Subscribe now

Subscribe now

I’m a marketer, borderline obsessive - so I tend to sweat the small stuff.

I think about when someone hears the ad, what are they likely to do next? Most will do nothing. This is the 95:5 rule. Even if they are my exact target, most are not sitting there with a live brief, a signed-off budget, and a burning desire to book a call with a man who has chosen to brand himself around foxes. Given this piece of empirical information, the first goal is to make sure at the very least that they remember my name or brand so they can find me when they do have a need.

So then my attention turns to the 5%. If they visit the site, what state are they in? If they are not ready today to contact me, what would genuinely be helpful to them? If they are ready, how do I make that step easy? If they do contact me, what would make the first conversation better?

If someone is not ready to talk, they may choose to read this newsletter. They can subscribe. If they do, I get a chance to stay top of mind. They may not want to subscribe. That’s also ok. They may choose to download the first two parts of my book as a PDF, ungated. No “work email required”. No lead-gen hostage situation where a perfectly nice PDF is held behind a little form like it has done something wrong.

For the people who are ready, when they visit my contact page, Isa, my chatbot assistant jumps into action and offers a 30-minute discovery chat. Isa’s job is small but does a lovely job bridging outbound marketing efforts (radio ad) to intake. It is also an opportunity to create a memorable impression for my brand. Never miss a chance to imbue the brand.

If they book a chat, the next bit happens in the background. The meeting goes into the diary. I get notified and another AI assistant starts researching the person and the company before we speak. So when the call happens, I can ask better questions earlier.

The system is a mix of things I built, connected to reasonably inexpensive tools - Vercel, Github, Cal.com, Substack and Healthcheck io. All held together by Claude Code. I focus on quality first - the experience, the branding, the value-add, my goals. Then I automate the parts that repeat.

That is what I mean by building systems. Designing the experience properly, then letting software carry the bits that should not need human effort every time. For me, the radio ad creates awareness and visitors. The website explains what I do. The book gives people depth without asking for anything back. The newsletter keeps the relationship open. Isa helps the ready people move. Cal.com removes the email tennis. The research assistant improves the first call. Each bit has a job.

I like this stuff because it is practical. It is not AI as theatre. It is not a robot doing jazz hands at a conference. It is a few small jobs, joined together, making a real business process less leaky.

Subscribe now

Jo runs client campaigns. This means it does all of this; list intake, audience and ICP selection, messaging, setup, launch checks, monitoring, paid support, reporting, and a bunch of boring rules that stop a campaign turning into a polite disaster.

I’ve built a few of these systems now. AI writers with messaging frameworks and voice rules. A qual researcher that can call your mobile and interview you. An AI ad engine that reads direct response craft files before writing a headline. Same pattern every time. Don’t give it a prompt. Give it a system. Documents it reads before it does any damage.

Subscribe now

When Jo gets a new campaign, it does not start by writing. It starts the way any decent marketer would. Who is the client? What do they sell? Who buys from them? What has worked before? What geography are we in? What channels are even allowed? Cold email in Ireland and the EU is not the same game as cold email in the US. Jo checks geography before suggesting a channel, which saves you from building a beautiful email sequence you can’t legally send.

Jo creates and manages what I keep calling a database. Technically it is a very large CSV file. Not glamorous. But works.

At the beginning of any session, it reads my notes from years spent in direct marketing. Bob Stone, Drayton Bird, John Caples, direct response writers who understood attention long before everyone started saying “performance” in meetings.

Before writing, Jo has to answer two questions. What is the strongest benefit to the reader? What would make a stranger stop scrolling? Only then does it write. It then goes through a slop audit. Em dashes, zero tolerance. Corporate language. Triple adjective stacks. Paragraphs with the same shape and length. Sentences that sound like they were written by someone trapped inside HubSpot trying to win a procurement tender.

If it reads like AI wrote it, Jo rewrites it before the client see it. Ironic I know. But not all writing is the same. Jo is not being asked to sit in a corner and have a point of view on culture. It is being trained for a narrow job: persuasive campaign writing, for a specific client, to a specific audience, inside a specific channel, with claims we can stand over.

AI has a confidence problem. Not emotionally, obviously. I don’t think Jo is lying awake worrying about its Q3 forecast. The issue is overconfidence. So I’ve trained it to treat claims differently. Some things are facts. Some are benchmarks. Some are hypotheses. I often download reports for Jo to read and upskill. But it knows that a vendor report saying “Campaigns that combine LinkedIn plus email produces 287% more responses” does not get treated as a fact. It might be true. It might also be a sales deck wearing a lab coat. Jo can use it as a hypothesis, but if we want to build a recommendation on it, we test it ourselves.

The same applies when results start wobbling. Jo does not immediately rewrite the copy just because reply rate is low. It checks the stage first: list, channel, deliverability, offer, message. Jo will not call a winner after 40 results because one version is slightly ahead and everyone is bored waiting. It will not call one bad week a trend when the list changed, or the copy changed, or it happens to be August and half the country has mentally left the building. Sometimes the answer is not “the campaign is broken”. Sometimes the answer is “too early”.

Over time, all of this turns into campaign memory. Every campaign adds a file. Every experiment adds a learning.

Jo can write a cold email, obviously. So can every other AI tool. The email is just the visible bit. The real work is the legal check before the channel recommendation, the craft file before the headline, the source rule before the claim, the diagnostic table before the rewrite, the campaign file before the analysis, and the learning log before the next brief.

That is what I mean by building systems. Not a chat window with a job title. A working system with enough process, craft, evidence, and memory inside it that the output has somewhere to come from.

The system still needs me for taste, judgement, and the times when Jo randomly decides to ignore the rules I sweated over to create. But for the most part, the repeatable checks, the nagging, and the boring bits that slip when people get busy now sit in the system.

Which is handy, because apparently I need managing.

Subscribe now

Brand marketer. Performance marketer. CRM specialist. SEO lead. Social strategist. Lifecycle manager. Content designer. Email automation expert. Analytics lead. Research lead.

It made sense when the tools were hard.

If you wanted a landing page, you needed a designer and a developer. If you wanted customer analysis, you needed an analyst with SQL access and a week of patience. If you wanted research synthesis, you needed a planner. If you wanted pricing analysis, you needed someone from finance who owed you a favour. I traded in coffee-favours as standard.

Subscribe now

So marketing became a relay race. Lots of handoffs. Lots of waiting. Lots of “can you put a ticket in for that.”

One marketer can now analyse customer interviews, synthesise research, build a prototype, create presentations, automate workflows, generate strategic options, analyse pricing models, interrogate datasets, write code, edit creative and produce decent execution, often without needing deep expertise in each discipline. Not perfectly. But good enough to make a lot of decisions, to ship stuff, to make some bets.

The advantage is no longer being the world’s best at one tiny thing. The advantage is being able to connect things together quickly. The marketers who thrive over the next decade will probably look less like narrow specialists and more like adaptable operators. People with a core strength, but broad capability across the whole commercial system.

That doesn’t mean expertise disappears. I spent most of my career as a brand marketer (if that is a thing). This is still what people usually contact me about - positioning, brand management, advertising effectiveness. But in the (recent past), that meant relying heavily on specialists around me. Designers, developers, analysts, researchers, production teams, strategists.

Now I can do things myself that would previously have required multiple people or agencies. Not because I’ve became an elite designer, developer or data scientist overnight, but because the tools lowered the barrier enough for me to do a lot of it myself.

Not everything. If you need a global brand re-launch with a multi-market campaign, I’m going to call in some of my expert specialist contacts to help this happen. But even for the big brands I managed - most of what we did were not these moments.

The people who may struggle most are the ones who define themselves too tightly by one function.

When your identity is attached to a role that’s shrinking, AI feels like a threat. When your identity is attached to learning and solving problems, it feels like a better set of tools.

A younger marketer entering the industry today has an even bigger opportunity. They can become dangerous across multiple domains very quickly. A few hours of focused learning with modern AI tools can compress weeks of confusion.

My advice? Stop seeing roles as fixed identities and start seeing them as capabilities we can learn. You probably won’t become the best data analyst, filmmaker or pricing strategist in the world. But you may become good enough to ask better questions, test ideas faster and execute without waiting for six departments and a Jira ticket.

The marketer who can connect research to strategy to execution to analysis without handing it off four times will just move faster than the ones that cannot.

Subscribe

The same structure every time. It hedges everything, fills gaps with fabricated confidence, and defaults to consultant-speak because it has nothing else to work from.

The problem isn’t creativity. It’s context.

AI doesn’t know what makes your brand different. It doesn’t know your tone, your audience, or what claims you can back up. So it guesses. And guessing produces copy that could belong to anyone in your category.

Professional copywriters don’t start with a blank page. They start with a brief, a brand guide, and a stack of source material. An AI writer needs the same things. The difference is you only need to build them once.

I’ve now built this system for several different brands. It’s not one oversized prompt. But rather a set of documents that together form an AI writer. Here’s what I build, and in what order.

Subscribe now

1. The messaging framework

Not a positioning document. Not a brand guide. One document that combines positioning, proof points, competitive contrast, voice rules, and hard bans.

I used to describe this as two separate steps: positioning and brand DNA. In practice, they live in the same file. The AI loads one document and gets the full picture: what we say, who we say it to, how we’re different, what the evidence is, and what we never do.

For one client, the messaging framework was 82 lines. One page. It had:

1. A value proposition split by use case, because the same company sells differently depending on which door the client walks through. Their general offer had one line. Their specialist offer had a different one.

2. Four messaging pillars, each with a “Proof:” line. Not “we’re experts” but specific evidence: years of experience, number of people, named platforms, client examples where previous attempts had failed before they rebuilt them.

3. A competitive contrast section. Not “we’re better than the competition” but a structural argument: “The big firms do X and hand you a manual. The small firms lack scale. We stay through adoption.” With a hard rule from the client: no negative shots at competitors. Position by what you do, not by what they do wrong.

4. A positioning statement in a specific format: “For [audience], [brand] is the [category alternative], combining [capability], so [benefit].”

5. And a “Never use” section and a “Never do” section. Words that were banned. Behaviours that were banned.

So one document. The AI loads it before writing anything, and every piece of content is anchored to the same pillars, the same proof, the same competitive framing.

For a larger company, the same architecture scales up. I ran a positioning workshop that mapped competitive alternatives by capability area. Not “who are your competitors” but “what would your customer use instead of you, for each specific thing you do.” Eight capability areas. Named competitors in each. Then three value pillars with specific proof points, best-fit and poor-fit account profiles, and an eight-step sales narrative that builds from market insight to the ask.

The positioning includes a story structure for selling. When the AI later writes a case study or an outreach email, it frames the problem the same way the sales team frames it. Without that alignment, the website says one thing, the pitch deck says another, and the AI picks up whichever it finds first.

The brand DNA sat alongside the positioning. Three pillars, each with explicit marketing implications. Not just names. Instructions. Such as: “Anchor every asset in a customer problem. Spell out their pain first, then show how we remove it.” Another said: “No safe B2B headlines. We prod with wit and clarity. Overcommit to one or two fights and throw our full weight behind them.”

The AI can write from “anchor every asset in a customer problem.” It struggles with things like “customer-centric.” One is a behaviour. The other is an adjective. Adjectives are decoration. Behaviours are instructions.

2. The voice framework

The messaging framework tells the robot what to say. The voice framework tells it how to sound, and importantly, how not to sound like AI.

These can live in the same document for a smaller brand. One client’s messaging framework had a voice summary and writing principles built in, all on the same page. For a bigger company with a marketing team, voice becomes its own document because the rules get long.

I currently divide my voice framework into three parts: voice traits with writing instructions, tone calibration by channel, and hard constraints with AI-specific detection rules.

Voice traits are not adjectives. “Infectiously upbeat” means nothing to an AI. I like it but robots not so much. They seem to prefer this kind of stuff: “Vary sentence length to create rhythm. Short ones inject energy. Longer ones carry the reader forward. Lead with positive framing. Active voice. Punchy verbs: ‘boost’ not ‘improve’, ‘kick off’ not ‘start’, ‘shift’ not ‘change’.”

Tone calibration matters because your voice on a campaign is not possibly not the exact same as your voice in a thought leadership report. The personality stays the same. But small adjustments. I’m the same talking regardless of if talking to my kids or my mates or my clients. But I curse less often (in front of my kids). I do this currently through a simple calibration table for each channel.

Then the hard constraints. I ban em dashes. AI uses them three to five times more often than human writers. They’re the clearest AI tell in writing. Zero tolerance. I ban corporate words. Unlock, leverage, ecosystem, synergy, reimagine, holistic, paradigm. If you wouldn’t say it to a customer over a coffee, it doesn’t belong. This is not so much AI but just bad writing. I’ve many, many words I ban. Delve, comprehensive, facilitate, utilize, landscape, embark, crucial, game-changing, revolutionary, breakthrough.

But bans on their own aren’t enough. AI doesn’t just fail at word level. It fails at rhythm level.

I build an AI Rhythm Detection section into every voice framework. It names the exact cadence patterns that give AI writing away:

• Stacked punchy sentences. AI defaults to: “Short sentence. Medium sentence. Marketing punch. Marketing punch.” Two punchy sentences in a row means rewrite.

• The three-point landing. Three consecutive sentences each ending on a “point.” This is the cadence of every product demo voiceover you’ve ever sat through. Rewrite immediately.

• The “This isn’t X, it’s Y” pattern. AI will use it four times if you let it.

• Uniform paragraph shapes. AI creates paragraphs that are all the same length and structure. Some paragraphs should be a single long sentence. Others should mix three different lengths. Predictability is the enemy.

I include a before/after comparison in each voice framework so the AI has a concrete standard. Here’s AI rhythm:

“The industry is changing. Fast. And the companies that adapt will thrive. The ones that don’t will be left behind.”

Here’s human writing:

“Teams know their pricing is a constraint, but their switching cost calculation keeps coming out negative. What is different is that customers can see this constraint now (unlike before), and this is making it more difficult to model.”

One difference to watch for; human writing develops an idea across sentences. AI writing restates the same idea with different emphasis.

3. Audience-specific messaging

One messaging framework is probably not enough if you sell to different audiences. The product is the same. But the message may not be. For one client, we built two separate positioning documents. Their two audiences face completely different competitive alternatives, different value pillars, different proof points, and different language entirely.

The AI writer identifies the audience before it writes a single word. It asks: “Who is this for? Audience A or Audience B?” Then it loads the correct positioning. An email to one segment uses one set of language. An email to the other uses different terms for the same capability. Same voice. Different positioning.

For another client, audience-specific messaging meant different tracks for different team members. One person targets leaders worried about a specific operational problem. Another targets executives going through broader organisational change. Same firm, same capability, but different competitive alternatives, different proof points, and different language.

We also include approved customer quotes per segment. Real words from real customers. When the AI writes for a specific audience, it pulls from the right quotes, the right pain points, the right proof.

4. Content-type frameworks

I’m coming to the view that the more narrow the use case the better. So my approach is evolving a bit. But for now, each content format gets its own framework. I build separate ones for advertising, email, thought leadership, case studies, and web pages. Each one teaches the AI why something works in that format, not just what to write.

A thought leadership report is not marketing with longer sentences. An email is not a landing page cut into paragraphs. Treating them the same is how brands end up with everything sounding identical and nothing performing. For example, for emails, we might focus on subject line formulas with performance benchmarks as well as body copy structures: PAS (Problem, Agitate, Solution), BAB (Before, After, Bridge) or AIDA.

The case study framework uses a five-part arc: company, challenge, solution, results, customer quote. Two intake paths: the AI asks structured questions in sequence, or works from whatever messy notes you paste in.Each of these frameworks is a standalone document, usually ten to twenty pages. They took time to build. But they only need building once.

5. The source protocol

This is what stops the AI from making things up. AI is confident by default. It will invent a plausible-sounding statistic (”73% of marketers agree...”) and present it as fact. Every writer who has used AI for more than a week has found a fabricated claim in their output. The source protocol makes verification mandatory, not optional.

Before writing anything, the AI completes a verification step. It identifies the audience. Reviews the positioning. Confirms what source materials it has: data, customer quotes, product specs, research. Every claim in the final copy must trace back to a source. No fabricated statistics. No invented customer quotes. No assumed industry data. If the AI can’t find a source, it flags the gap and asks.

I also build in credibility stop triggers. If the AI is about to use a percentage without a source, a performance claim without documentation, or a company reference without permission, it stops. It doesn’t invent. It asks.

Every response the AI produces ends with a protocol confirmation block: which audience, which documents reviewed, which voice settings applied, what materials used, and what gaps found. If the block is missing, the output is incomplete. Don’t deliver incomplete work. The gap between AI writing and professional writing isn’t creativity. It’s credibility.

6. The orchestrator

The master instruction file ties everything together. It tells the AI who it is, how to route requests, and what workflow to follow every time.

When a writing request comes in, the AI follows a decision tree: identify the audience first, then the content type, then load the correct positioning document, voice framework, and content-type framework. An email to one audience uses different positioning than a thought leadership piece for another, even though both are for the same brand.

The workflow is always the same: identify the audience, load the positioning, confirm the source protocol, write. Quality checklist at the end. If anything is missing, the AI doesn’t deliver. It asks.

For one client, the orchestrator is 280 lines. It routes between two audiences and four content types, enforces the voice framework on every response including conversational replies (not just deliverables), and requires the protocol confirmation block at the end of every output.

What you end up with

Once the documents are in place, the AI writes from your positioning, in your voice, for your specific audiences, in whatever format you need. Case studies, email sequences, thought leadership, web pages. Different structures, different tones, same brand.

For one client, the system is five documents. For another, eight. Both took a few days for the first pass, then refinement as teams used them, feedback and we fine tune them. Testing them takes longer than building them.

The voice framework requires the messaging framework, because you can’t define how you sound until you know what you’re saying. The content-type frameworks require the voice framework, because each format is a variation on the same voice. The source protocol requires the content frameworks, because each format has different credibility requirements.

Skip the messaging framework and every piece of content is generic. Skip the voice rules and everything sounds like AI. Skip the source protocol and the AI starts inventing claims you’ll have to answer for.

So this is how I built AI writers. Slowly. Then fast.

Subscribe now

We place bets. We just don’t call them bets.

We should. In a world where we now get fast, easy and confident answers, it is a good habit to learn. Instead of asking if this is right or wrong, the question is more like ‘what kind of bet is this?’

Subscribe now

I think in bets. I didn't always but it is a habit for me now.

It’s not about the maths. Mostly. It’s so I can think more clearly without tricking my brain. If you say ‘this is going to work’, you’re committed. 100% committed. And it is hard to find flaws. Now your brain needs it to be right. Consistency and commitment kicks in.

Over time, I’ve trained myself to give the bet a number. Not a precise number, not a spreadsheet number, just something like “I think there’s a 60% chance this might work.” The difference between that sentence and “This will work” is huge, because the second one locks you in and the first one lets you change your mind.

Someone says “I’d put it lower, maybe 40,” and instead of a fight you’ve got a conversation about why. You can disagree with a probability without it feeling personal. Try disagreeing with someone who’s just said “I know this works.” That’s a different meeting.

Not all bets are equal, either. Spending an afternoon testing a new channel is a cheap bet. If it doesn’t work, you’ve lost an afternoon. Committing half your budget to a platform you’ve never used is a different kind of bet entirely, and it deserves a different level of confidence. Some bets are low risk. Some bets are high risk, but maybe worth it.

The language of bets lets you say what’s actually true: this one is low risk and reversible, that one isn’t, and we should probably talk about that before we commit.

Annie Duke wrote a whole book on this, Thinking in Bets. She was a professional poker player. I did a stint in online poker. I wasn’t any good but I was surrounded by smart folks, many of whom had played serious poker. I suspect that poker pros naturally think about everything in bets.

The other thing is that a good decision can still lose and a bad decision can still win. If we only judge our calls by what happened afterwards, we learn nothing useful about whether the thinking was right.

Philip Tetlock found something similar in Super Forecasting. He discovered that the best forecasters weren’t smarter or better informed. They were just less married to being right. They updated when the evidence changed, instead of defending what they’d already said.

He called them foxes.

Subscribe now

We get about 2-3 seconds before people scroll past or stop watching our ad. I’m generalising in the sense that some ads might get six seconds in the right channel, with the right creative. But we have this tiny window of seconds, where we must to get our brand into people’s memories.

They must notice our brand if we want them to retrieve it later. Not just notice the ad, but the brand.

Subscribe now

Plastering your brand name from the opening frame is an option but it can feel bolted-on or makes it harder to land your idea. That’s why Distinctive Brand Assets are so useful. They do the heavy lifting.

In 2021, when I was running marketing at The National Lottery in Ireland, we would show this waterslide image in our research. No name. No logo. No tagline. Yet 72% of the country could still spontaneously name the brand.

The waterslides were a DBA that we had built through consistent and rigorous advertising. People recognised the brand without us having to cram our name into opening frames. This gave us freedom to push creativity, safe in knowledge that every ad impression wasn’t falling at the first hurdle - getting the brand noticed and hopefully into long term memories. People knew who was advertising so we could focus on making the ads enjoyable - another useful goal in ads.

We can credit the brilliant Jenni Romaniuk with the phrase, and the science behind Distinctive Brand Assets. I wrote about them in my first book six years ago. And later when I joined The National Lottery, we built a bunch of Distinctive Brand Assets and were extremely rigorous and disciplined in using them.

We built them, and then sweated them it in all our communications. Our logo, our tagline, our waterslides, our green background and our music track - we used consistently as a toolbox. And we measured them often. You cannot manage what you don’t measure.

This first campaign, internally known as “William” was our first introduction of the asset, and our re-introduction of the tagline “It could be you” which we’d had dropped some years earlier. Few brands had taglines that were liked, famous and still relevant. So it was an easy enough decision to bring it back.

The following year we launch Village, which played into a different insight but allowed us to tell new, entertaining and joyful stories linked to our waterslides. Joy is a big thing for this brand.

We resisted the temptation to move away from this, and instead always looked to see how we could build onto it, to have ‘fresh consistency’. For example, we focused on the yellow inflatable ring and developed that into a distinctive asset itself - which we measured.

Character Brand Assets

Character assets are some of the strongest Distinctive Brand Assets companies can have. One advantage of character brand assets is flexibility. You can use them as heroes in your ads, not just aesthetics within the ads. Over time, if consistently used, people recognise your brand at a glance. This is far more important than folks realise.

We created a Dream Inspector character for this reason, and built this into a recognisable asset. Within a year or so, over 60% of adults could attribute our brand to this character.

You can see how we used the Dream Inspector into our short format activation ads in the opening seconds.

Use AI to sweat your assets

The challenge in past was production. I could do a two-day shoot and plan in as much footage, photography as possible to get all sorts of ads for all our channels. I was lucky that we had the resources for this.

Small teams and small brands do not. But even with that, it had to be planned ahead. Few brands have writing teams and art directors ready to create new ads or content every week. Not if you care about quality and consistency. You can argue they don’t need to, but this is a different discussion.

AI removes that constraint. If you are willing to built a system.

Grumpy foxes

The title of my first book and now my consultancy is “Run with Foxes”. I’ll explain why in another article. But my grumpy fox is my Distinctive Brand Asset character. Or foxes - there’s more than one.

I’ve built an ad engine that creates ads using my fox. I can brief my AI system and get back something in a few minutes. The level of consistency, right down the grumpy attitude of the fox is impressive.

It took a few marathon sessions to get this right but once built, a lazy prompt like ‘grumpy fox drinking coffee’ will get me what I want almost always first time. If it makes a mistake, my focus is not on the output, but on why it didn’t do what I hoped, and how to fix for next time, so it works every time.

We started with images, but I’ve patiently trained it to understand that images need space for headlines. It must think about headline placement before it creates the images. It has been trained for contrast, for different fonts styles, different colours, alignment of the headline - all the small decisions you would intentionally make when you are making ads.

It can create video too. I’m still just creating short videos, under 10 seconds. But that is mostly what is needed for ads. I’ve focused on two types of videos. One is my tarantino-style cinematic ones. So high on visual impact, low on any messaging.

But often we need video ads that are more informative, with datapoints, benefit headlines or charts, CTAs. For these, the role of the fox is just a small nudge to make it easy for people to notice the brand. The headline is often the hero. These are what some call “sales activation” ads.

My focus is quality and speed. Speed is easy if you don’t care about quality. But I do. Getting both takes more time.

I find the best way to approach this is to build each system for a very narrow use case and put the hours into getting that exactly right, every time. For example, I’ve lost hours of my life testing this over and over to make sure the headline is not touching the fox. I’m not saying this is a good way to spend my time, but it upsets me otherwise.

How it writes headlines

I spent many years in direct-response advertising, the predecessor of performance marketing. Back in my youth, I earned the finer details of what headlines work and why. I studied the classics - Drayton Bird, David Ogilvy, John Caples, Claude Hopkins, Bob Stone.

So before writing any headline, I’ve trained my system to answer two questions from classic direct response advertising: what is the strongest benefit to the reader? And what would make a stranger stop scrolling? Only after answering those does it write headlines.

Build a system

The fundamentals are the same. Attention decay is a thing. Most purchases come from retrieving stuff from our brains - our memories are the biggest search engines. We must make sure the brand gets noticed, not just the ad. And building Distinctive Brand Assets is as important as ever. What’s changed is that one person with a terminal window can sweat these assets every day, in every format.

My focus is not making an ad. But building a system that makes quality ads consistently, easily.

Subscribe now

It's a Tarantino trunk shot, straight out of Pulp Fiction. Desert sun behind him. He looks like he's deciding what to do with you.

The whole thing runs through Claude Code, the AI coding assistant, orchestrating calls to external image and video generation APIs. No Photoshop, no After Effects, no video editing software. Just a terminal window.

Subscribe now

The character

The fox is a custom AI character. Not pulled from a library. He was designed through iterative prompt engineering with specific physical traits locked in: big golden-yellow eyes (always heavy-lidded, never wide), fluffy orange fur, white chest patch, dark brown mitten paws, oversized head with chibi proportions. Pixar-style 3D rendering.

The personality constraint is critical: he is bored and unimpressed, never angry, never scared, never happy. Even in danger, he looks annoyed. The expression reads differently depending on lighting. Bright, warm settings make him look grumpy. Dark settings make the same face look sinister. So every prompt has to manage that.

A set of 16 training images were created to establish the character across angles and expressions (front face, waist shots, full body, smirk, surprise). These serve as reference material for the image generation model.

Step 1: Generate the still image (Seedream 4.5 on Replicate)

The image generation uses ByteDance’s Seedream 4.5 model, hosted on Replicate. Claude Code writes and executes a Python script that sends a reference image of the fox plus a detailed text prompt to the API.

The prompt for the Tarantino shot was:

“Classic Quentin Tarantino trunk shot from Pulp Fiction. The grumpy fox stands at the back of an open car, looking down into the trunk. He wears a dark charcoal suit, white dress shirt, black skinny tie. Both paws rest on the trunk rim. The shot is photographed from INSIDE the trunk looking up at the fox. The interior walls and carpet of the trunk are visible in the foreground at the bottom of the frame. The fox’s body and face fill the upper half. Blue sky and blinding desert sun behind him. His expression is cold, menacing, unamused. Wide angle lens distortion. Low camera angle from below chin level. Dusty desert highway, New Mexico. 35mm anamorphic film, heavy grain, warm amber tint, slightly overexposed highlights. Gritty 1990s crime film aesthetic.”

That level of detail matters. Rich prompts beat generic ones. “Weary resignation of someone who lost a bet” produces better results than “grumpy face.”

The API call costs about $0.03 and takes roughly 15 seconds. Claude Code polls the Replicate API until the image is ready, downloads it, and opens it for review. Nothing moves to video until the image is approved.

If Seedream fails after a couple of attempts (it sometimes puts literal cameras in the scene when you describe camera positions, for example), there’s a fallback: a custom LoRA model trained on those 16 fox images, hosted on Replicate.

Step 2: Animate the image (Kling 2.6 Pro on fal.ai)

Once the still image is approved, Claude Code uploads it to fal.ai‘s storage and submits it to Kling 2.6 Pro, an image-to-video model. This produces a 5-second video clip.

The video prompt is minimal compared to the image prompt. The model can see the image, so you only describe motion. The rules are strict: one motion arc per clip, a mandatory personality enforcement sentence in every prompt (”The fox maintains his grumpy, irritated facial expression throughout”), and contact points locked down (”Feet do not move or lift”).

For the Tarantino shot specifically, Kling’s start-and-end-frame feature was used. A code-generated near-black frame served as the start image (Seedream can’t render darkness, so a black frame was created programmatically), and the fox-looking-down-into-trunk image was the end frame. Kling animated the transition between them, creating the trunk-opening reveal.

This step costs about $0.35 per clip and takes a few minutes.

Step 3: Assemble the finished ad (Python, locally)

Claude Code runs a Python script using PyAV and PIL to composite the final video. A standard fox ad is 7 seconds, 210 frames at 30fps.

All of this is handled programmatically. No editing software involved.

The Tarantino clip was generated in both landscape (1920x1080) and portrait (1080x1920) formats, because the homepage uses responsive video: desktop visitors see the landscape version, mobile visitors see a portrait version trimmed to 2.8 seconds.

What Claude Code actually does

This is the part that surprises people. Claude Code isn’t just writing the scripts. It’s orchestrating the entire workflow in a single terminal session:

1. Writes the Seedream prompt based on the creative brief

2. Generates and runs the Python script that calls the Replicate API

3. Polls the API until the image is ready, downloads it, opens it for review

4. Waits for human (my) approval before proceeding (this is a hard rule)

5. Writes and runs the Kling animation script via fal.ai

6. Polls that API until the video is rendered

7. Writes and runs the assembly script that composites the final ad

8. Manages file naming, aspect ratios, and format decisions

I tell Claude Code what I want (”make a Tarantino trunk shot”), it writes the prompts, calls the APIs, shows you the results, iterates if needed, and assembles the final output. The documentation, the code, and the creative decisions all live in the same conversation.

There are currently nine approved ads in the archive. The Tightrope, The Dab, The Haaland Meditation, The Head on Desk, The Walk Across, The Hathaway Man, Coffee in Soho, and two versions of the Tarantino trunk shot (portrait and landscape). Plus a 48-sheet billboard of the Hathaway Man. Five of those nine were made in a single Saturday afternoon.

The cost

A finished ad costs under 50 cents in API calls. The main investment is the iteration time: getting the prompt right, reviewing the image, deciding whether the motion works.

The website implementation

The finished Tarantino video sits as a full-bleed hero on runwithfoxes.com, which is a Next.js site. Two <video> elements with css media queries handle the responsive swap between landscape (desktop) and portrait (mobile). Both autoplay muted, play once, and hold on the last frame. Poster images provide the static fallback before the video loads.

The whole system, from character design to image generation to video animation to ad assembly to website deployment, runs through Claude Code talking to three external APIs.

We didn’t build a video. We built a system. The fox character, the creative pipeline, and the production workflow are all documented in a single guide that any new Claude Code session can pick up and use immediately.

I really like the end result. While we have a system, it is not flawless. But I fancy the challenge.

Subscribe now

Sounds odd but it is the result of something I’ve been tinkering away at for a bit. I wanted to see if my AI team could do what good teams do - work, debate and ideally agree on priorities, without me there to drive the discussion.

Subscribe now

I’ve changed client details, but this is what minutes from their meetings look like when I go check:

To backtrack a bit - Klara, Jo and Nat are agents. My agents. A project manager, a growth lead, and a finance one.

They each run on their own role file with their own brief and their own stubbornness, and every morning after the briefing pipeline finishes, they sit down with each other and argue about what the top seven on my briefing page should be. I insisted I couldn’t handle more than seven.

So when I come back to the kitchen, the list is there. A JSON file with a trade history attached.

They each share priorities from the things they are tasked to do. For example, Klara wanted a whole bunch of items to put in front of me, all tasks for a big outbound client launch thirty days away. It wanted me to review, get my act together and sort out the email domains, go over a walk-through plan for the messages. It also had written feedback it believed I owe a client, which it spotted in a meeting transcript from a couple days ago.

Jo pushed back. It had two replies sitting in my gmail inbox from warm contacts, both landed on Thursday, both now at day three of the seventy-two hour window it treats as sacred. Growth execs take these windows very seriously, even the robot ones.

Apparently my average time to reply is about four days, thus the perceived urgency. And it wanted Klara’s feedback item off the list, because I talk to that client nearly every day, so it’s not a thing to worry about.

Nat doesn’t care about replies, ever. Nat cares about invoices. It puts the biggest unpaid line in the book on the (virtual) table and reminds everyone that Paul is shocking at getting invoices out. On average I take about twenty days to get around to sending one, and Nat has the data to prove it. That is fair. I’ve never been a big fan of admin tasks - thus perhaps my enthusiasm for building these systems.

Klara is the decision-maker within this mighty AI team. I generally don’t play favourites among the robots but somebody needs to do it, and in fairness, Klara is a project manager. Says it in its role description. So it decided what gets paused and what gets into my top 7. Nat’s two unpaid invoices get pushed to tomorrow, because nobody knows how old they are. Including me I suspect.

The final list goes to my /today briefing page with the trade log attached.

It all sounds like fiction, and I’m not going to claim three agents arguing produces more accurate answers than one well-prompted call. But I get my list of 7 things to focus on today. That said, writing this wasn’t on it.

Subscribe now

I knew ‘deploy’ from war films. Troops being deployed. Serious men pointing at maps. I did not associate it with me, in my kitchen, at half ten on a Wednesday, pushing a website live while the kettle boils.

Subscribe now

The progression crept up. First, Claude helped me write things. Copy, paste, tidy it up, done. That was genuinely useful and also, in hindsight, the shallow end of the pool. Then it started creating actual files. Not suggestions, not drafts in a chat window, but documents I could open, save, send.

Then we started building software together. I describe what I wanted, watch it appear, click through it, show it to someone. I’d built an event system in an evening. A dashboard that runs my business. Not because I suddenly understood engineering, but because the cost of trying had dropped so far.

That’s where the vocabulary changed. I stopped saying “I made a thing” and started saying “I deployed it.” The word just showed up, like it had been waiting in the hallway. I didn’t decide to start talking like a developer.

Soon I’ll have delusions that I can judge the quality of code like I do the quality of Effie Award entries.

But while I can deploy, I cannot yet make things secure. I cannot guarantee they won’t break when real traffic arrives, or quietly leak something they shouldn’t, or fall over at 2am when nobody’s watching. Nearly half of all AI-generated code contains security vulnerabilities, according to Veracode. I believe it. I’m writing code I can’t read and pushing it live to people who trust that someone checked.

That gap, between “I can deploy” and “I can keep it running,” is where the next conversation lives. And it’s a conversation we haven’t had yet, not properly, because the relationship between marketers and IT was never designed for this.

The old model went like this. Marketer writes a spec. Spec joins the queue. The queue has a queue. Somewhere in a building you’ve never visited, a team you’ve never met adds your request to a backlog that already has forty-seven items on it, three of which are from people who’ve since left the company.

Six weeks later you get a meeting invite to discuss the requirements for the meeting that will discuss the build. By the time the thing ships, the campaign it was meant to support has been live for a month, the brief has changed twice, and the original stakeholder is on maternity leave.

Both sides slightly annoyed, both sides slightly justified. I sat in those meetings for years. In fairness, so did everyone.

That model assumed marketers couldn’t build. The spec was the contribution and the building was someone else’s job. The whole workflow, the briefs, the tickets, the prioritisation ceremonies, the quarterly roadmaps, existed because marketers needed permission to make things.

We don’t need permission anymore. We can prototype in minutes, click through journeys, show people real things instead of describing real things. Why write a document about the thing when you can build the thing?

But deploying is not yet the same as building.

Building is the prototype on your screen. Deploying is the URL in someone else’s browser. And once it’s in someone else’s browser, you need the things you never thought about. Security. Reliability. Scalability. All the words that IT teams have been quietly handling while marketers were in the other room complaining about the wait.

So the conversation changes. It used to be “can you make this for me?” Now it’s closer to “I made this, can you help me make sure it doesn’t fall over?” The marketer is showing up with working software instead of a PowerPoint deck. The IT team is being asked to harden something that already exists instead of building something from a brief. The marketer feels like they’re being helpful. The engineer feels like they’ve just been handed a live grenade wrapped in a nice font.

I suspect the marketers who do well here won’t be the ones who just learn to deploy. That part is getting easier by the month. We’ll soon go from ‘I’m working on a deck’ to ‘Here’s a working prototype’.

It’ll be the ones who learn to properly get things built, quickly, securely. The ones that sit down with someone from engineering or security and say, honestly, I built this and I don’t know what I don’t know. Which, granted, is not a sentence they teach you in marketing.

I still say “deploy” twenty-five times a day. It’s genuinely feels close to the finish line. But maybe it’s still closer to the starting pistol.

Subscribe now

Subscribe now

Thursday, March 26. An “updated” timestamp. A little block of copy that starts with “Two meetings that matter today” like it’s my job to be organised. You can see that Claude even reminded me about the proposal due March 23, still open. Slightly judgy in my opinion.

I didn’t add a task for that proposal. Claude knows because it read it in a meeting transcript. When I type /today, Claude scans my emails, my calendars, my meeting transcripts.

The dashboard has a task page. I can add tasks the normal way. But Claude also proposes tasks it thinks I’ve committed to, and it gives me three options, Accept, Edit, Dismiss, like I’m moderating my own life. If I hit Accept, it goes into my list. If I hit Edit, I can fix the title or the date before it becomes real.

It’s not a nicer task list. It’s a system that keeps comparing what I’ve said I’d do to what I’ve actually committed to somewhere. If there’s a mismatch, it doesn’t create tasks behind my back. It politely proposes one, tells me why, and waits for a button press.

Then there’s my Progress page. I am very proud of this one. If you want to know what matters to somebody, look at what they do, not what they say. Claude tags my meetings, my emails, and the work I do inside Claude, which is most of it, and it tracks the time and the tokens without me having to label anything. That’s the trick. If I have to tag stuff, I won’t. Or I’ll do it for two days, then lie to myself.

So now I can see the week as it actually happened. I spend a lot of my time building things, which is a priority for me at the moment. I think it should be a priority for all marketers.

I built the dashboard, partly out of curiosity of what it is possible (a lot) but also to help me from doing admin badly, or doing it twice, or doing it late. I’ll write about it in more detail in my book.

Subscribe now

Subscribe now

You’ve seen it. Every stat is an animated counter. Every testimonial fades in word by word, as if the customer is currently remembering their own quote. Parallax text, scroll-triggered reveals, typewriter effects. Then there’s the hover-reveal card. You move the mouse and it flips, expands, slides, hands you the “extra detail”.

These patterns are seductive. I’m guilty of them. I genuinely like how they look.

On their own they’re aesthetically pleasing. This isn’t the situation like all the horrible purple AI gradients on every tech site in 2024. These sites look good. Professional.

The problem is they look the same.

Tom Roach at BBH wrote about this a few years back. “Sameness is commercial suicide”. He wasn’t talking about scroll animations. He was talking about brands, and how brand copy each other to the extent that they look look the same.

Or sound the same.

And once sameness becomes the default, you stop noticing it. You just feel vaguely bored and move on. Tom was writing about brands more generally, but websites are where a lot of brands now do most of their talking. Especially in B2B world. If your site looks and behaves like everyone else’s, you’re not just copying a layout. You’re stepping into the same bucket of indistinctness and hoping your logo will do the heavy lifting.

I saw a thing just yesterday from a guy called Paul Bakaus. He wrote a set of design rules called Impeccable, with one simple test, would someone look at this and immediately assume AI made it. Not whether it’s ugly. Whether it’s generic. Sameness. Same interaction patterns, same animation density, same assumptions about how people use the web. The tools are converging on a shared aesthetic that nobody sat down and designed, and it’s spreading because it looks good enough not to argue with.

You don’t fix this by banning animation or swearing off AI tools. You fix it by choosing, properly choosing, what you want the thing to be, and then building around that instead of around whatever defaults happen to ship well in a generator.

Subscribe now

My mate will spend most of St Patrick’s Week (yes it is a week) working on a presentation about their brand architecture. Not the brand architecture itself. The presentation about it. The deck needs to look good. He’s spent weeks on it already. Weeks on wrapping paper.

Subscribe now

I’ve been there. Many nights I’ll never get back dragging text boxes around PowerPoint, picking a font weight that didn’t make me wince, and moving a logo three pixels to the right because it was annoying me. Nobody in the audience would notice. I do unfortunately.

If you manage a marketing team, do the maths on how many hours per week your people spend formatting slides. Not thinking about what to say, just making it look acceptable. It might be the single biggest time thief in marketing that nobody talks about, because it feels like work. It looks like work. It even sounds like work when someone says “I’m working on the deck.” But it’s not the work.

The wonderful news is we no longer need to do this. You can build a system inside Claude Code that takes any document, a Word file, a PDF, a set of bullet points, and converts it into a branded, scrolling webpage. Navigation, animations, the lot.

I tested it on a few articles and presentations. One prompt. I waited four minutes. I did not move a single text box.

This is not a rough prototype. The aesthetics are as good as your instructions. My brand guidelines are at runwithfoxes.com/brand so Claude follows these carefully. They look like this.

This is not just about visuals. It is thinking about the experience. When people talk about taste, this is what they’re talking about.

So I personally care about craft - so things like the frosted-glass navigation bar that starts transparent and gently blurs as you scroll down. I like the sections that alternate between a light background with a subtle dot-grid pattern and deep charcoal panels with cream text. I want the pull quotes where the orange border grows from nothing to full height, then the words fade in.

Data gets turned into Chart.js visualisations, bar charts that grow, doughnut charts that fill, all triggered by scrolling. Hover over a card and the background shifts. Key quotes type themselves out letter by letter as you reach them. There’s a thin orange progress bar across the top of the page that fills as you read, so you always know where you are.

None of it is decorative for the sake of it. A static slide deck is something you click through. This is something you scroll through, and the difference is that scrolling feels like reading, not waiting.

I tried this claude skills on an old presentation I had on my laptop from the wonderful folk in Ebiquity, called “Profit Ability”. I asked Claude to look at it in my files and use my skill to recreate this in my branding. Just one prompt. These are screenshots, but these charts have lovely subtle animations too.

What Claude Code is actually doing is worth understanding, because this is the pattern that applies to far more than presentations.

I wrote a “skill”, which is a set of instructions that Claude Code follows every time I trigger it. The skill tells it how to read the source material, how to map the content into sections, which fonts and colours to use, where to place animations, how to build the charts, and what the final page should look like. It’s a recipe.

The first time took a while to get right, because I cared about the details, the exact hex values, the type scale, the spacing, the way the nav should behave on mobile. But once the recipe exists, it runs the same way every time. Claude Code reads the document, writes the HTML, the CSS, and the JavaScript in a single file, copies in the brand assets, and opens it in the browser. One command. The same craft, repeated perfectly, without the craft taking any time.

That’s the key idea for teams. You don’t lower your standards. You encode them.

The fussiest designer on your team, the one who notices the three-pixel gap, they’re exactly the right person to build the recipe. They do the hard work once. Then everyone else gets their taste for free.

Now, there’s a version of this story where I tell you the tech is the point. It’s not. The point is what it frees up. When you can turn your thinking into something shareable in minutes instead of hours, you stop hoarding your work until it’s “ready.” You share earlier. You share rougher. You get reactions from stakeholders while the thinking is still alive, not after you’ve spent so long on the slides that you’re emotionally attached to the structure.

We’ve all been in the meeting where someone presents a deck they’ve clearly spent days on, and the feedback is “actually, can we go back to slide three, I think the whole direction might be wrong.”

The hard work is the thinking. The formatting should be free.

If the gap between knowing what you want to say and having it in a format people will engage with is where your team’s time goes, that’s not a skills gap. It’s a tools problem. And your mate shouldn’t still be dragging text boxes on Paddy’s day.

Subscribe now