Promptly Transparent

It’s clear that there are a ton of ways to use an LLM (ChatGPT, et all). They are therapists, authors, co-authors, research partners, planners, explorers, answer machines, and an infinite array of other things. Overall, they are the dawn of the conversational interface.

Computing is moving off the desktop and off the web. LLMs are not the cause, but they are the accelerators. Both desktop and web computing have been trying to break out of these containers for more than a decade. They tried (and mostly succeeded) with mobile. But that kept the core computing metaphor intact. They have crept into your car, your kitchen, your headphones, your front door, your heating system, your bedroom. The interface stayed more or less the same.

It's been a challenge to shake loose of the broadcast media paradigm. Interfaces (from newspapers, radio, and TV to their contemporary expressions) have been built based on one-to-many communications (broadcast) models. They’ve depended on scale to get their messages across. That idea (that one interface is sufficient to broadcast the same message to an audience) is very sticky.

It’s an industrial era mindset. One factory, many customers is the heart of the idea. Until now, that meant a limited array of possible options. Industrial thinking never really got all that far from “You can have any color as long as it’s black”. It spread gracefully into SaaS software.

Always, the industrial Idea is a one-to-many deal. That way, you can consolidate investment. We’ve taken the ‘many to many’ path in social media, gaming, online work.

LLMs usher in the possibility of 1 to 1 computing. The business model is evolving rapidly. Though I’m certain that the goal is to assemble a large audience, what I’m seeing is an astonishing level of personalization (I hesitate to call it intimacy). The technology meets its users at their level of need. Each individual user interface is a complex interaction that tells you who the other person is.

I sit in a seat that allows me to see the ways that intense users approach getting value from LLMs. It’s idiosyncratic. Each user expresses their personality, beliefs, concerns, and appetites in their prompts.

Recently, a good friend and I went about trying to understand MCP (It’s a protocol created by Anthropic that is fast becoming the LLM equivalent of an API). My answer was a 40-page document that included code snippets. His answer had an executive focus. Mine took a full day of work with Claude. His was a quick stroll through the park.

His prompt felt airy to me. My prompt was more of a process of chasing an answer by discovering pieces of it. I wanted to know the whole story. He had extremely specific questions. What I learned from his prompt told me a lot about what matters to him. Reading it made getting aligned with him that much easier. Who he is shone through the prompt.

There is real power in sharing your prompts and your approach to using these new tools. Right now, it’s something we do in private. There are no standards (yet). We are all groping around in the dark with relative degrees of intensity.

I am going to start sharing my prompts, my process, and my results from time to time.

The most recent dive into a rabbit hole took two full days of thought (and battling with the limits of $20 accounts).I set out to try to understand how all of the many forms of AI could be applied to HR. (LLM is a small piece of the puzzle).

My notion was that by understanding the tools and the totality of the set of use cases, I could start to really see how the future of work unfolds.I used multiple LLMs. They all deliver different opinions based on their training data and the way it was ingested.

Basically, as I hit one account limit in one, I moved to a different tool until I hit its ceiling. Here is the path I took:

1. Define and understand the various AI tools. (Depending on how you count and which tool you ask, , there are between 15 and 30).
2. Get a grip on which tools are deterministic (will always produce the same answer) and which are probabilistic (won’t always produce the same answer)
3. Find 7 to 10 use cases in HR where a particular tool could be applied (it started out as 3, grew to five, then to 7, and finally to 10.
4. Take those 250 use cases one at a time and evaluate the quality of their output (expressed as a % likelihood of a ‘correct’ answer)
5. Identify use cases that are beyond the reach of an AI solution. (These turn out to be situations where elevated levels of uncertainty are the coin of the realm (Like disciplinary work and CEO pay negotiations). There were about 100.
6. Build a spreadsheet. (I’ll share it in the comments).

Each step of the process was a flurry of progressively better prompts. I came to understand my question fully as a core part of this research. The result was an interesting map of the possible ways to introduce AI into HR.

(A good friend notes that when you inject AI into HR, you get HAIR.)

By understanding the breadth of HR possibilities, I got a good look at the edges of the HR domain. That’s where the real transformation is going to happen.

I’d love to see your prompts and/or understand your usage methods.

I don’t think there is any wrong way to use your unique interface. Prompts are clear expressions of the things that drive your curiosity. That’s something I want to know about you.

It’s also a way to promote increasingly sophisticated usage of the tools. Please share your prompts. We can learn a lot from each other.