Antonym: The Cognitive Engineer Edition

The best results are less about prompt engineering than thinking about thinking.

Dear Reader

Imagine if there was a map of everything we know about AI.

"What we know" is a fast-expanding island, roughly circular, that grows daily as practice and discovery open up capabilities and methods. It's like those diagrams of continental drift, but faster.

The frontier, the edge of what we understand, has a circumference that is increasing at an exponential rate. For now, it's called AI literacy, and everyone is able to exchange ideas and techniques no matter their field.

One day the borders of the newfound knowledge will become so stretched that it won't be useful anymore to think of it as a single thing. The island will break into an archipelago of specialisms – still close enough to hop from one type of knowledge to another. The core island will be for beginners, a foundational set of principles that everyone must learn, like persuasive writing, calculus or the scientific method.

This fragmentation is already happening, perhaps too soon to be easily allowed. Several times recently, I've heard people talk about "a pause" in generative AI developments, which they naturally welcome as a chance to draw breath. I think that’s a wishful delusion, perhaps a necessary one – a story conjured by minds willing it to be true so they can catch up. Looking across different fields and industries, the reality is that progress is at best keeping the pace of the last two years, if not accelerating.

The Great Connection

I kid you not, I was about to cancel Notion AI because we were using it so little compared to other LLMs and tools. And then they literally upgraded it the next day. Now Notion connects to our Google tools – Gmail, Drive, Docs, Calendar – and actually understands them and can:

• Have a guess at timesheet entries

• Prepare you for a meeting by finding notes from the last one

• Take a look at all the actions noted for you and help you prioritise

• And it can look at... (drumroll) ...Slack! Which is the equivalent, if you're a Microsoft Office user, of it being able to see all your Teams chats, as well as Outlook, OneDrive, etc.

So can Claude. In the same week, they added the ability to look across our drives and email. It is incredible. And this is all before you get to the even more clever ways that technically inclined users can use MCP (Model Context Protocol) to connect to all sorts of things.

Paddlers Have Opinions, Practitioners Have Insights

There are two types of content you'll come across in places where people talk about AI: from paddlers and practitioners. The paddlers have trod in the shallows of AI and found it too cold to go further. It seems they want to convert everyone else into understanding this reluctance as rational.

Here are some insights from practitioners that I've noticed:

• The massacre of the darlings: Software engineers are used to killing their precious code when something better comes along. Writers can learn from this.

• Absolute forkers: Developers save versions of their work and create alternates to see what works. This approach helps avoid the paralysis of trying to get it perfect the first time.

• It's all data: Seeing your writing as malleable data rather than precious art changes your relationship with creation.

• Getting it out of your head: One of the things that defeats us with large projects is the cognitive load of holding concepts, craft, knowledge and other details in our heads while writing. Going from 50 to 50,000 words comes with exponential complexity. AI helps offload this burden.

My Recursive Self-Replication Challenge

For all the worrying about AI replacing our jobs, the most requested, most useful and most tricky thing people want to do with AI is replicate themselves. Either their role, or the boring bits, or the bits they find hard – like writing.

My struggle to replicate my writing tone of voice through AI has gone through these stages or experiments:

1. Simple prompting: “Write like this” – and attaching an example file – doesn’t get close enough.

2. Custom GPTs with examples of my writing: One of the first custom GPTs I made was Mayfbot, trained to write like me. It kind of did, but not reliably enough to use. Like the single prompt, the analysis of style was too shallow. I’m irreverent at times, but the AI would make me sound like a snarky blogger. Imagine….

3. Poe bots with swappable models: I was kind of happy with Poe for a while. On this platform you can create custom bots and then swap out the AI that’s running them – Claude instead of ChatGPT for instance – and tweak the instructions easily. I like the bots I made with other people’s voices (George Orwell, Margaret Atwood were among the experiments), but my own never seemed right. You know your own voice. Or do you? Maybe hearing an AI version of yourself can be as uncomfortable as hearing a recording of your own voice, with some uncanny valley thrown in. With a recording you think, “do I really sound like that?”, whereas with AI you think, “Nah - I definitely don’t sound like that.”

4. Claude Projects (which are really good): ChatGPT rival Claude doesn’t offer the ability to create the equivalent of custom GPTs on its platform, but it has rather good Projects (although the amount of data you can add to the knowledge base is weirdly limited compared to GPTs). Claude 3.7, indeed all the Claude versions so far, is preferred by many, including myself for writing prose. It’s always more… human than its rivals.

5. And now: Recursive prompt engineering: where we ask the AI to produce detailed prompts based on analysing my style. My latest crop of bots and Projects that can write prose or rewrite it to sound like me are getting much better. The leap has been driven less by the advances in AI models as how I think about tuning and improving prompts and custom GPTs. The not-a-secret trick is this: asking the AI to improve your prompts or help you build them. Then having another conversation or system critique the outputs and add them with your own notes to creating the bot or process again.

It isn't just the tech that changes – it's us. We're adapting our approaches as the tools evolve.

The best results are less about prompt engineering than thinking about thinking. Metacognition. Cognitive engineering, then – designing the processes and steps we will put our minds and their machine partners through to get a thing done.

Things I'm Trying Out

Using AI to make software (vibe coding)

Look! I made some software that makes diagrams. I kind of love it. Both the fact that I made software (I can’t write code) and that it does exactly what I want.

It took around an hour and a half to get it to this state. While far from perfect, it’s exactly what I want for a specific job: bringing processes to life in presentations and documents.

I used a vibe-coding platform called Bolt to make Diagrammar, which lets you publish it to share and use right away. I’ve put this one in my browser’s bookmark bar and used it to make diagrams like the one above this article.

What I learned was:

• Start with a clear idea of what your app needs to do. Having something you can’t do with your existing software but really want to is a great place to start

• Use AI to help write the brief. I developed a detailed description of a simple app to solve my problem in ChatGPT and then brought it to Bolt’s AI chat.

• Add one simple feature at a time. Keeping the complexity low helps build a better app.

• When you get stuck, roll back to a previous version before you started to try and build the feature.

• If it all feels horribly wrong, ask the AI to do a retrospective, listing what it learned in the process and what it would do differently next time. Copy this and add to your brief or use as a brief and start again.

Treating prose like programming

This is really interesting, and something I will return to in future Antonym. Reading about and playing with vibe-coding platforms has meant inadvertently learning more about how software developers work than I expected to.

One of the most important AI tools for coders is Cursor, an app that’s a lot like the ones they have used for creating software, but now with AI features to speed things up.

Cursor's allows you to write like coders code – I’m not sure if it was designed for this – where prompts sit in one folder, text in another, and research in a third. This allows for a new kind of creative flow where OKRs expand into briefs, into PRDs, into analysis.

Bonus fact: Cursor was valued at $9 billion this week in a funding round.

Cursor’s parent company Anysphere closed a US $900 million growth round on 5 May 2025, led by Thrive Capital with Andreessen Horowitz and Accel following. The deal values the two-year-old AI coding-assistant maker at ~US $9 billion, up nearly 4× from January’s US $2.5 billion mark. Reported annual recurring revenue has already passed US $200 million, so investors are effectively paying c. 45× ARR for a still-private software business.

Sources:
Financial Times
CryptopolitanCrunchbase
NewsSiliconANGLE

That's all for this week.

Thank you for reading. If you liked it, stick a 'like' on it and I'll come up with something again next week.

Happy vibing,

Antony