Magic

Magic and Indranet.

Kidding, of course, but... perhaps less than you might think. It can certainly seem like sorcery, can't it? You put in "write me one o' dem dere customer person things" and a few seconds later, you get:

                 Please assist in developing a detailed customer persona. Include the following key attributes:
                 1. Background: Brief life story, career, and lifestyle.
                 2. Demographics: Age, location, education, and income level.
                 3. Goals: Primary and secondary objectives this persona is trying to achieve.
                 4. Challenges: Main obstacles preventing them from achieving their goals.
                 5. Product/Service Relationship: How our product/service can help overcome their challenges and achieve their goals.
                 Feel free to ask for specific details or examples to ensure the persona aligns with our target customer segment.
            

So, how did we DO that? It's a lot of pieces, if you look real close, but it's a pretty simple idea overall.

First, we teach the model how to be especially good at a given task.

Marketing, creative writing, designing mnemonics, coding in Rust, sounding like Donald Trump; these are skills - areas of narrowed focus and competence - with which we can tune a given LLM session into resonance. We can get it to stop thinking about rainfall in Nigeria and the history of the Algonquin language group when it should be writing a saucy greeting card for your husband. A very large part of that amounts to teaching the model how to understand user intent - to grasp what it is you actually want. This is key - the art of expressing yourself in a way that the model can understand well is a large part of the skill of prompting. A good prompter can write something that contains the needed context to guide the model to your desired answer. They prompt well so you don't have to. A very good prompter can write something that gives you not just what you ask for, but also what you actually need.

We are extremely good prompters, indeed.

Second, we show it a large number of examples of exceedingly varied and particularly efficacious prompts.

A "Greatest Hits Gallery", effectively. We show it a small sample of the wide variety of prompting available, and the different sorts of approaches that may be taken. This is vital, as the model has been trained on only the most rudimentary, primitive, caveman-level prompting up to this point. It simply has never seen much good prompting before. If it starts "Act as a..." or has a name like "Tree of Auto-Freshness", it's almost always trash. This is why simply telling the model "Write me a prompt that does X" produces such predictably, aggressively mediocre results without significant guidance. The model is great at prompt tactics - "rewrite this rambling mess into something prompty" - but is simply terrible at advanced prompt architecture or novel structures and notation.

Third, we teach it how to plan and think about strategies and goals.

This equips it to actually figure out what it should do before it starts doing it. Usually it's more of a "I'll build my wings after falling off the cliff" sort of thing.

Finally, we regularize the prompt.

We use a fairly rigid framework to express the final prompt. This usually will be of a fairly regular structure of a list of detailed instructions prefaced by an objective and some context. We preface it with an "Effectuate the below:" to obviate the model from saying "That's a great fremework! If you follow it closely, you're sure to succeed!" or similar, which can happen, otherwise.

And what lets it all happen is the connection to Indranet.

Indranet is Collaborative Dynamics proprietary AI orchestration layer. It's an agentic environment that allows one to trivially connect arbitrary AI models to arbitrary internet resources with programmatically precise flow control, fidelity, and data management. This lets us very, very easily connect a series of LLM model-instances working in concert precisely as we wish, instructed as we wish, marrying the accuracy and control of computers with the judgment and creativity of AI, harnessing it all to perfect understanding of human intent.

It is not an agent framework. There are not independent programmatic entities acting independently. It is an agentic network - it gives agentic power to the entire field of AI as a whole, en masse, all at once, to any who choose to use it. It is collaborative. It is very dynamic.

It is Indranet and we believe it to be the future.