WordPress is experimenting with an AI generated “podcast” feature allowing you to select a handful of posts to be turned into a conversation. Novelty notwithstanding – not bad, but still a long way to go. You decide.
Pip: Alan Williamson has been quietly dismantling the mystique around AI, one plainly-worded post at a time — and apparently also building fake corporate executives to fight each other in Slack.
Mara: That is a fair summary of where we are headed. The posts this episode cover how AI prompts actually work under the hood, and what happens to workplace culture when people stop thinking and start forwarding.
Pip: Let’s start with the mechanics — what an agent actually is, and why the architecture diagrams are doing more work than the code.
AI prompts: simpler than the pitch decks suggest
Mara: The central claim in “It’s Just a Series of Prompts” is that the scaffolding sold as sophisticated AI infrastructure is, once you strip the branding, three things: skills, tools, and memory.
Pip: And the post is direct about what those actually are: “Skills are markdown files. Plain text. Instructions written in English that tell the AI what to do in a given situation. Not code. Not configuration files with obscure syntax. A text file you could write in Notepad.”
Mara: So the upshot is that anyone who can document a process already has the hard part covered. The rest is telling the AI which file to read next — which is roughly what the post demonstrates with a working commute-check skill built from a single text file.
Pip: The buzzword layer is doing a specific job, and the post names it plainly: it is pushing you back, making you feel like implementation requires specialists and budget sign-off. The complexity is a sales posture, not a technical reality.
Mara: “Behind the Prompt Curtain” goes a level deeper and explains the stateless mechanics underneath all of this. Every message you send bundles the entire conversation history and ships it to the model fresh. There is no memory on the server side — the chat client is maintaining the illusion.
Pip: Which means the conversation array you manage yourself is the memory. You own it. That is also why long sessions drift, why a new chat window loses context, and why the context window has a hard ceiling.
Mara: The post also clarifies how tools work — the AI does not run anything directly. It returns a structured request asking your code to run a function, and your code sends the result back as another message. As the post puts it: “The AI is the decision maker. You are the hands.”
Pip: That framing matters because it is also where cost accumulates invisibly. Every tool definition you include consumes tokens on every request, whether or not that tool is relevant to what the agent is doing right now.
Mara: The post flags this specifically for popular frameworks like Hermes and OpenClaw, which ship extensive tool libraries loaded by default. The advice is to configure which tools load per context — most frameworks support it, and most implementations do not bother.
Pip: From architecture diagrams to token budgets — and the next question is what all this efficiency machinery is doing to the humans running it.
When output replaces thinking
Mara: “AI is Making Us Artificially Intelligent” opens with a business proposal that arrived with emoji headings — a rocket ship, a lightbulb, a key next to the key takeaways — and the observation that nobody had read it before sending it.
Pip: The tell is always the emoji. Genuinely, though, the post’s concern is structural: volume has become a proxy for effort, and AI has industrialized that instinct at scale.
Mara: The numbers in the post are stark. AI-generated content crossed fifty percent of new web articles in November 2024, up from five percent before ChatGPT. Meanwhile, average human screen attention has dropped to 47 seconds. The post introduces the acronym BOVINE — Being Overly Verbose In Needless Expositions — as the name for what replaced KISS.
Pip: And then there is “Simulating Corporate Chaos: Your Own Truman Show,” which takes the cultural observation somewhere unexpected — building a multi-agent simulation of C-suite executives in a Slack channel, each powered by a different model, capable of web searches, memes, and private one-on-one conversations about their colleagues.
Mara: The project is open on GitHub. The point is that the same mechanics powering the agent framework post — memory, tools, context management — are enough to produce behavior that reads as recognizably human, including the clichés.
Pip: The through-line across all of this is that the complexity was never in the technology — it was always in knowing what to ask, and whether the answer was worth asking for.
Mara: That question does not get easier when the tools get cheaper. Next episode, we will see where that thread leads.
Posts the podcast refers
- Behind the Prompt Curtain
- AI is Making Us Artificially Intelligent
- It’s Just a Series of Prompts
- Simulating Corporate Chaos: Your Own Truman Show
AI Disclaimer: Gemini Nano Banana Pro was used to generate the photo – 2012 Pitch Perfect, and WordPress the AI podcast
Alan Williamson 