A common observation is that it’s not using real fonts. An AI image model makes a bitmap image of what it thinks a poster should look like. That has image has shapes that resemble letters, spelling out words. But unlike the words you’re reading out now, those shapes aren’t built from a library of letter shapes (or more precisely, strict recipes for producing shapes) called a font. It’s as if the image model is hand-drawing each letter from its memory of what that kind of typeface looks like.

Sometimes that’s obviously appropriate. Psychedelic Sixties and Push Pin Studios clearly rely on letter forms that a real designer would hand-draw.

Other times it appears more likely that a real font would be the correct choice. Swiss / International Typographic Style would almost certainly use Helvetica, Contemporary Editorial looks like a mainstream serif font like Times Roman. In this case are we only using the model’s approximation because it doesn’t have another way?

A designer friend observes: “I don’t recognise any of the fonts, they’re all approximations. One previously unseen font on a poster I could accept, but 2-3 when they’re all fairly bland in any case still registers as a tell for me.”

As I’ve pointed out before, I don’t care about tells: fooling people that they’re not looking at AI is not a goal. But not triggering icky feelings, is a goal. If some people are getting an uncanny-valley response because that’s not quite Helvetica, this is a problem. (Notwithstanding, the same people get not-quite-Helvetica ick from Arial).

So I thought I’d experiment by trying to convert some posters into a vector format. Cut out the graphical elements, find a font that matches the text, and reassemble it into an SVG file. The story of how that went, and all the results, is on its own page.

When it works, you get pretty close matches, and the knowledge that the type is as a type designer intended. Here’s the De Stijl gig, before and after:

So that’s better, right?

Maybe sometimes, maybe not always, I’m not sure.

Look at the Bauhaus / Modernist gig, below. That pointy ‘V’ the model drew works really well for me. Perhaps if my convertor had found the right font, the conversion would have a pointy V too. But that might not match other letters. And anyway, the point is that the AI model is free to decide that here, in this one word, in this one position, a pointy V works.

In other places, like the Festival of Britain gig below, the model-generated drawn text lines up just-so: “CARVER &” the same width as “THE FLOOD”. You can achieve this with type by stretching or adjusting whitespace, but the model has achieved it in its drawing with quite a sophisticated combination of both: drawing and spacing all the letters so that in combination they fit the space.

And on that theme: one of the reasons to use fonts with a real text layout engine, is that the layout engine knows how to space letters just-so. And that’s especially ideal when you’re typesetting blocks of text. But I’ve been learning that especially with big titles and logos, it’s not unusual for a designer to nudge things around. You don’t get that for free.

The AI model has the freedom to:

• invent letter forms that deviate from what the “font” would normally use
• use different weights, widths, spacings, shapes according to context and space needs

… indeed it has no concept of not doing those things.

So myself, I’ve become quite attached to letting the model do this. A model with suitable training could become very, very good at it – I’m just using general purpose models, though. Who knows whether OpenAI are training an image model in that direction?

I should mention the other reason to build posters in a vector format: editability. If you have a poster in a combination of bitmap objects and vector elements, you can move things around, add or change words, and so on. That is not the focus of this blog post.

As ever, I am not a professional designer, nor even a well trained amateur. I can only speak for my own untrained eye.

before	after

Expressionist Woodcut, the same fête, version 1 and version 2.

Back in June I wrote about why AI posters don’t have to be horrible and allowed my experiments to sprawl into a website with a hundred poster prompts. I was pretty pleased with them.

But having these examples made me fussier, and I began to notice flaws.

Then, after essentially nobody had paid any attention to the page, I shared it on Hacker News - where lots of people commented, many of whom noticed the same flaws as I had:

The text layout was usually pretty good - it would emphasise the headline, and put other details where they belonged - but it could have been better.

The fete example highlighted a common AI problem - if a word is in the prompt, it’ll put it in the image. So the image for the fete at the cricket ground for the village church restoration fund, would contain cricket, a church, fete tents, as well as imagery representing every attraction listed.

(By the way, you’d get a better poster by just listing fewer attractions – but our hypothetical client wants them all on the poster, so that would be cheating)

This was just one contributor to clutter. The model often seems to want to fill every space with stuff.

The fictional band highlighted another problem: AI had invented “Carver and the Flood” as the band, playing a gig at “The Anchor”, so many of the posters featured a woodworking tool, a lot of water, and an anchor. Sometimes these were lovely images. But it’s very literal - and for a real band, “The Flood” would have ceased to be associated with water, in the fans’ minds, years ago.

The ferrets. My fictional fete had ferret racing as an attraction. So many of the posters highlight ferrets. This is understandable - it’s the most interesting thing about the event, so the model picks it out. But you don’t want it every time. You want the user to have the opportunity to pick a try where the ferret isn’t the star.

Often the model would invent logos. So the church repair fund suddenly has a crest. That’s not acceptable.

How version 2 was built

I put all these issues to the Claude chatbot, and it suggested a number of ways to improve the prompt. This went back and forth quite a lot, because it would come up with rules that work for some styles but not others.

Eventually I had it write a handover, so I could move into Claude Code.

When I wrote v1, after assembling the code to produce the prompts, I spent a couple of evenings on the part-developed site, copy-pasting prompts into ChatGPT, saving the results to my project, as fast as the rate-limiter would allow me.

This time I paid for some OpenAI API credit, and automated it. All I intended to do was make the generation less laborious.

But as soon as Claude Code realised I had the API, it suggested doing an evaluation exercise. So that’s what we did.

Producing an image is two steps: the first is prompt expansion. We submit the prompt to a LLM, and it generates the text explaining what it’s going to do. You can review that before sending it to the image model.

So, to avoid spending too much money on image generation, we evaluated that first step. I was able to leave a lot of this to Claude Code. The criteria were pretty clear, so Claude could be the judge, Claude adjust the prompts, Claude could decide when to stop.

while results aren't quite right {
   amend prompt
   submit prompts
   judge results
}

I just had to keep an eye on spend. And to my surprise it was iterating by making hundreds of requests, but costing me pennies. But we were having to firm up the prompt to tell the model things it should have worked out. Claude had picked five styles to test against, one for each way a poster can decide what to draw, plus a deliberately awkward one:

• Polish Poster School, where the event chooses the subject, so the ferret problem shows up here first
• Art Nouveau, where the tradition insists on its own subject — a woman among flowers — whatever the event says
• Suprematism, which is abstract, and should draw no subject at all
• Factory Records, where the typography is the image
• Madhubani, which fills every square centimetre, and would tell us whether the new rules about restraint had flattened a style that is supposed to be dense

Then we realised we were testing with gpt-5.4-mini - very cheap, but not very clever. Switching up to gpt-5.6-sol – what ChatGPT would use on Medium – started behaving much better, and we could remove many of the adjustments we’d made trying to make the mini model work.

All this meant that with about $10 worth of tokens, we’d evaluated and improved our prompts until they met the goals.

The biggest sticking point was ferrets. I wanted it to choose a ferret about 20% of the time. It’s a striking image, many users would choose it, but we want it to use other options when you retry. The problem was that it’s unusual so the LLM finds it compelling.

Where the subject comes from

For some styles, the event dictates the subject. A fete in Letterpress, gets some fete-related image.

For other styles, the style dictates the subject. Art Nouveau is a woman with flowers. You can only override that by asking.

Other styles are explicitly abstract.

Cost

This cost me $28 in OpenAI tokens. Around $10 in text-only evaluation runs, then another $18 doing multi-run image tests on four styles, followed by generating all 200 images.

Oh well.

Outcome

I think in general the posters are better. They’re more striking. The hierarchy of text is clearer. The subject matter is wrong less often.

You still need to tweak the prompt, or just roll the dice a few times. The website shows the first shot for each prompt, warts and all. A second try might have been better, or might have been worse.

The new prompts are way longer than V1, and perhaps take some control away from the user. Used as intended - copy-pasted into ChatGPT - it can of course be tweaked. But the audience that would tweak it is possibly an audience that doesn’t need the hand-holding these prompts do.

Ferrets?

It still chooses to use ferrets a lot of the time. Well, the ferrets are good. But, by design now, in ChatGPT it’s going to show the user what it’s chosen, before producing the image. It offers a shortlist of other subjects, and the user can ask for any of the alternatives.

Other stuff

It’s been pointed out that you might be able to get better results by letting AI generate a code-based format like Postscript or SVG. Here it could lay out real fonts, give you elements you could adjust yourself. This is all true, and something that’s either already available in tools like Photoshop or Canva, or with a skill and a code agent.

But here we’re trying to demonstrate what a harried PTA parent can achieve in a chatbot.

One hundred poster prompts.

Maybe I have poor taste, but I enjoy both the process and the outcomes - and not having authored them directly myself perhaps gives me a more detached position from which to decide that I think they’re good.

But I have done a lot of thinking about what I have and haven’t done. At first I thought of myself as the author, fooling myself into thinking that I had provided the inspiration, and the AI had merely saved some of the perspiration. Looking back at the logs, though, at times I was surprised at how many ideas had emerged from the AI, and at others I had forgotten what I myself had supplied or rejected or shaped.

I perceive a new landscape of issues, where humans collaborate with AI for creative works, among them:

• A widespread visceral distaste for anything AI has touched.
  That is, of course, allowed. But my own opinion is that this fails to recognise the good material AI can generate, while also failing to account for effects of the strong guiding hand of a human co-author, commissioning editor, director, producer – whatever you want to call it.
• Humans seeking to pass off AI-assisted creations as their own work.
  I feel this is both unethical and wasteful. It’s deceitful not to acknowledge AI’s role in the creation. While it’s often a good idea to clean up distracting AI tics from the language, it is wasted effort to try and deceive the reader into believing this is purely human writing.
• A missing experience for readers.
  It seems that readers (and viewers of art) like to imagine the author at work. When they read a sentence, they are thinking about what lived experience caused the author to use those words. Perhaps they imagine the author speaking directly to them, as they read. Their assumptions may be false, but they are nonetheless part of the experience of reading. Believing that the entire text fell out of a language model in a matter of seconds, disrupts that experience. Learning later, that something you had imagined a human author writing, came from a machine, is the unpleasant experience of being conned. Yet when a piece is co-authored by a human and AI, there is a writing process that the reader may think about as they read.
• Readers don’t know what they’re getting.
  Without an honest, supplied provenance, how does a reader know whether they’re reading something wholly written by a human, wholly generated by an algorithm, or produced with some level of human curation by a human/machine centaur? That context can be important to the reader.

If your framing is where mine is, it can be thrilling to know that AI, rather than the human, came up with certain ideas. In Crab, it is the model that notices that letters scattered on the floor can be misconstrued as “EVAC”, and works that into the story. Amazing.

For these reasons I decided to adopt radical transparency. I’m reluctant to even call myself an author, though I still don’t know what word to adopt. Since I don’t really know what to call what I do, I just demonstrate it - you get near as dammit a complete transcript of the interactions that built the piece. My workings.

You can read any of my stories without reading the workings - but they are there to skim or read in depth, as an honest record of how much I did, and how much a language model did for me.

Then I thought I’d better explain myself. And I thought, who knows, perhaps other people would like to adopt this practice? So I wrote a manifesto. Not entirely as a joke.

Read the manifesto here: The Kladde 26 Manifesto: Show Your Workings

I’ve taken inspiration from the Dogme 95 manifesto, hence the Danish name, though Kladde’s aims are quite different from Dogme’s “vow of chastity”. Dogme constrains what a director may create. Kladde allows you to create what you like, how you like – but you must disclose your methods. Kladde refers to the rough-book Danish schoolchildren use for preliminary rough-workings. Of course the Kladde manifesto comes with its own workings.

Lars von Trier didn’t stick to Dogme for all his films, and (at time of writing) one of my stories lacks workings. I’m trying.

If you want to join in, you are free to just get on with it. You can scrape the manifesto, link to it, or you can adopt the practices without referring to the document. I don’t care. If there’s enough interest I might even step up to a domain name. (update - I got that domain name!)

The workings of this post? I wrote it all by myself (though I admit to asking Claude for comments).

One big thing is that we can now produce code that’s probably fine, but has to be reviewed and tested, faster than we can do that reviewing and testing. How do you deal with that?

My favourite LLM and I wrote an essay together, with animated diagrams, even, explaining how one of the solutions to this is backpressure.

Or in plain language, slow down and stop swamping me.

Read the full essay →

A now famous Facebook post shows us the scourge of identikit posters generated by AI.

Here’s an article on the subject from the Independent.

Here’s another I found in the wild. With apologies for picking on the Leamington Beer Festival - they are by no means unique

The problem with these is not so much that they’re bad. They’re OK, I don’t love them. The problem is that once you’ve seen that style 20 times it starts to irritate just from the sheer repetition.

I knew that even ChatGPT was capable of a broader variety of styles than this, so I set out to prove it.

I gave ChatGPT some invented event details and asked it to produce a poster. Rather than go with the default look it would choose for a spring fayre, I made a point of specifying what I thought was a different style:

Me:

Produce a poster for a spring fayre.
21 April - 11am to 3pm
Mill Beach Park, Honeyford
Free entry
Tombola
Cakes and drinks
Performance by a samba band and a dhol band.
Craft stalls
Circus Skills workshop

Go for a clean, unfussy, bright layout with a bold striking spring-themed graphic. Avoid pastel/airbrush/oil style art or images of people.

This is what came back.

Not as successful as I hoped. To me it still feels like what I’m trying to avoid.

I asked for something deliberately different.

Me:

Make another one using a completely different design aesthetic of your choice. Treat the current one as a “what not to do” – not that there is anything wrong with it, but we want ours to stand out from other posters that look similar.

I think this is already much better. It stands out just be being different. I wondered what I would call this style if I wanted to ask for it again?

Me: What is this design style you’ve used called?

ChatGPT: This one is leaning into a modernist / Bauhaus-influenced graphic style, with a bit of contemporary “flat illustration” layered on top.

More specifically, you could describe it as:

Bauhaus / Modernist Poster Design

• Rooted in the Bauhaus tradition (1920s–30s)
• Strong emphasis on function over decoration
• Asymmetrical layout with clear hierarchy
• Bold, simple sans-serif typography
• Limited, high-contrast colour palette

Geometric Minimalism

• Built from basic shapes (circles, semicircles, rectangles)
• The flower is abstracted into almost symbolic forms
• Very little texture or shading — mostly flat colour

Swiss Style (International Typographic Style) influence

• Clean grid-based alignment (even if loosely applied)
• Information-first layout
• Clear, legible structure

If you wanted a shorthand label for a designer or printer, you could call it:

“Bauhaus-inspired geometric minimalist poster”

If you’re trying to stand out, this direction works well because most local event posters default to:

• bunting
• hand-drawn florals
• pastel palettes
• “craft fair” aesthetics

Whereas this feels more like:

• a gallery poster
• a design museum flyer
• or something you’d see in a well-curated café

If you want, I can push it further into a really distinctive niche (e.g. brutalist, risograph, 90s rave flyer, Japanese minimal, etc.).

Well, thanks GPT. I knew there was an endless array of design styles that it could produce. Instead of trying to dredge some names from my memory, I asked for suggestions.

It replied with a list of styles, some of which I’d never heard of, some I recognised but couldn’t picture, some which were names I didn’t know for styles I did.

Me: Give me a list of styles to choose from

ChatGPT: Here’s a deliberately varied menu—some safe-but-distinct, some quite left-field. I’ve kept each one concrete so you can imagine how your fayre poster would actually look.

🎨 Clean but Characterful

1. Bauhaus / Modernist
(What you just saw)
Geometric shapes, primary colours, strong type
Feels cultural / design-led rather than “craft fair”

2. Swiss Style
strict grid, lots of whitespace
Black/white + one accent colour
Very “information design”, almost institutional

3. Contemporary Editorial (Magazine-style)
Big headline, elegant serif + sans mix
Think weekend supplement or gallery leaflet
Could feel surprisingly premium for a fayre

🌿 Graphic & Illustrative (but not twee)

4. Risograph Print Style
Limited ink colours, slightly misregistered layers
Grainy textures, bold shapes
Feels handmade but designed, not crafty

5. Cut Paper / Collage (Matisse-inspired)
Inspired by Henri Matisse cut-outs
Organic shapes, bold colour blocks
Very striking from a distance

6. Botanical Scientific Illustration (Modernised)
Clean line drawings of plants, labelled or diagram-like
Monochrome + one highlight colour
Quietly distinctive, slightly academic

⚡ Bold / Unusual Directions

7. Brutalist Graphic Design
Raw, high-contrast, almost “ugly on purpose”
Heavy black blocks, stark type
Feels rebellious—very different from typical village posters

8. 90s Rave Flyer / Acid Graphics
Neon colours, warped type, energetic layout
Could echo the samba/dhol energy
Surprisingly effective if done with restraint

9. Memphis Design (80s Postmodern)
Playful shapes, squiggles, clashing colours
Inspired by Memphis Group
Fun but still graphic and bold

🧭 Quietly Distinct / Understated

10. Japanese Minimal Poster
Very restrained layout, lots of empty space
One strong graphic element
Elegant, calm, and unusual in this context

11. Monochrome + Single Accent (Ultra-minimal)
Black, white, and e.g. bright green or orange
Typography-led, almost no illustration
Relies on composition rather than decoration

12. Wayfinding / Signage Style
Looks like a park map or transport signage
Icons, arrows, structured layout
Functional but visually interesting

🎪 Slightly Playful but Still Clean

13. Modern Icon System
Everything represented as simple icons
Consistent stroke weight, grid-based
Feels like an app UI or transport system

14. Stamp / Letterpress Style (Modernised)
Bold inked shapes, slight imperfections
Limited palette, strong contrast
Suggests “event” without being twee

15. Festival Poster (Contemporary Indie)
Big central graphic + stacked lineup text
Borrow cues from music festivals
Works nicely given the bands

If you want a steer

Given your brief (“clean, unfussy, bold, no twee imagery”), the strongest stand-out options are:

• Risograph
• Cut paper / Matisse
• Brutalist (toned down slightly)
• Japanese minimal
• Wayfinding / signage

If one of these clicks, I can generate a poster in that exact style—and push it properly, not just superficially.

Asking for specific aesthetics produced a variety of posters.

Now, I’m not claiming these don’t have a whiff of AI about them. I don’t think that’s really my goal - we can be honest that we’re using AI, we just don’t want it to have that look that everyone’s sick of.

Me: Please give me Stamp / Letterpress Style (Modernised)

Me: and now please Japanese Minimal Poster

Me: Now do Memphis Design please

(I didn’t know what Memphis Group design was, but I recognised if when I saw it.

Then I got a bit silly. Designers Republic is a design agency famous for record covers from the late 80s onwards.

Me: And now in the style of Designers Republic

One note here is that along the way, the poster has acquired extra text (“A day of music making and family fun”) – it’s decided that Designers Republic would have thrown in some text of that kind, so it’s come up with some words. Those words are now in the chat context, so they appear in subsequent posters.

If you were working from scratch, you’d just ask for the style you want in the first place, and avoid picking up trappings as you go.

Me: Now as if a professional graphic designer has added typography to a poster paint drawing by his young child.

Me: Now with the aesthetic of a photocopied 1980 punk fanzine. But with some colour snuck in.

Me: Now please make one in the style of a 90s drum n bass gig flyer, with early 3D/fractal computer imagery

Back to sensible ideas, naming an art style but asking for a poster that movement might have used, seems like a strong tactic.

Me: Now in the style of a contemporary 40s poster for a cubist exhibition.

So there we are. The sharp-eyed are still going to recognise these as AI-generated, but they’re distinctive, and I think they’re not ugly.

So that’s the moral - you don’t have to make posters that look like everyone else’s.

There’s a further step - you don’t have to ask AI for posters as images that can’t be edited. Claude and Gemini can generate HTML, PNG, PDF, where the text is real text, layers are real layers, you can change fonts, move things around, edit the text – but that’s more than I want to talk about today.

Inspired by this blog post, I went on to make a catalogue of one hundred poster styles, with ready-to-paste prompts and example images for each one.