The Agentic Approach

Fluid: Forward Compatibility of Prompts

Written by Iggy | Dec 10, 2024 12:36:10 PM

Will your prompts perform in the next version of ChatGPT?

Background

Last Week, Zackarias (our Director of AI) and I were deep in a session refining our internal AI implementations. I noticed something peculiar—each time we structured a rule or imposed a limit, I’d twitch. 

It’s not the first time this has happened. 

This is also what sparked Inspectrum, an AI-framework going towards interpretation and understanding of vectors rather than just standard similarity search. This holds the key to a completely new way of allowing for conclusions without metadata allowing us to work ourselves down to a lower level AI-language, this creates both efficiency as well as interpretability. The idea started just because I can’t stand structured data, in relation to AI it gives me the ick.

Actually, structured data gives me the ick in general, there is something obscene and unnatural about it.

What Zackarias and I unearthed during that session was a pattern I’ve been following unconsciously in every app I’ve built with AI—a refusal to instruct the AI and avoidance of all structured data. Instead, I provide context and goals, trusting it to navigate the space in between. When needed the goals need to be better explained and the context expanded. 

This isn’t just intuition—it’s a necessity. In a world where large language models (LLMs) are advancing faster than we can predict, writing rigid instructions assuming current functionality ties us to the present. A prompt, in its most beautiful form, should transcend time, being able to guide through direction instead of control, enabling handling of any unknown capabilities that tomorrow’s AI will bring.

The Philosophy of Fluid Prompting

Traditional programming is built on instructions: do this, don’t do that. It’s deterministic, precise, and, for most systems, entirely appropriate. But it also assumes that you know what the system can or cannot do. And AI isn’t deterministic—it’s probabilistic, heuristic, and dynamic. To harness its true potential, we need a new paradigm: fluid prompting.

Fluid prompting doesn’t tell AI what to do—it invites it to explore possibilities within a defined context and toward a clear goal. It’s less a set of rules and more a shared vision, a way of saying, “Here’s what we’re trying to achieve. This is the overarching goal. This is the expected output we want.” Allowing the paths to improve and find new paths between the beacons of goals, as the AI’s grow stronger and more capable.

This approach isn’t just about flexibility—it’s about ensuring that what we build today remains compatible with tomorrow. A rigid prompt might work perfectly with today’s model, but it risks becoming obsolete when the next generation emerges. A fluid prompt, by contrast, is timeless. It’s semantically rich and clear, capable of adapting to capabilities we can’t yet imagine.

The Beauty of Timeless Prompts

I believe the most elegant prompts written today will one day be seen as classics. They will hold intrinsic value, not just for their functionality but for their semantic perfection. Like a perfectly balanced equation or a beautifully written algorithm, a timeless prompt is one that remains clear and effective no matter what unknowns are thrown at it. This is what we strive to achieve in our platform to ensure that we can manage it all on a minimal amount of people, our solutions growing with the AI effortlessly.

 

Samples from our session:
Building an Action:

  • Rigid Prompt: "Never exceed 300 lines of code when building an action."
  • Fluid Prompt: "Ensure that the action remains atomic—a reusable, smallest possible function."


The rigid prompt imposes an arbitrary limit, tethering the AI to a constraint that may be unnecessary or counterproductive. It might even encourage the AI to focus on getting close to 300 lines instead of as few as possible. On top of that it is always tough using numbers and Math in LLM’s, its like talking the lowest level language to the highest level runtime or interpretor. The fluid prompt, on the other hand, focuses on the essence of the task—atomicity. It’s a principle, not a rule, and one that can evolve with the system.

 

Constructing an Agent:

    • Rigid Prompt: "An agent should never consist of more than 10 actions."
    • Fluid Prompt: "Optimize for efficiency. Determine how many agents and actions are required to accomplish the task at hand."
  1.  
  2. Again, the rigid prompt limits the AI’s creativity, assuming a static environment. The fluid prompt trusts the AI to decide the optimal structure based on the context, freeing it to adapt as capabilities expand.

Gathering API Information:

    • Rigid Prompt: "Always ask the user for the API manual."
    • Fluid Prompt: "Ensure you have the latest updates and information about the API in question."

    The rigid prompt assumes that real-time API access isn’t possible—a valid assumption today, but one that could become irrelevant tomorrow. The fluid prompt, by contrast, is future-compatible, allowing the AI to adapt as its capabilities grow. This isn’t the cleanest, but it reduces the impact immensely.

Forward Compatibility Through Semantical Motivation

Inconsistency isn’t just about contradictions within an agent or action—it’s about contradictions in their ability to remain relevant as the technology evolves. A rigid prompt might work today but fail spectacularly tomorrow. Fluid prompts, with their focus on goals and context rather than instructions, are inherently forward-compatible.

The trick is to craft prompts that hold no contradictions—prompts that are as functional and meaningful tomorrow as they are today. When done right, a prompt becomes a kind of art form, a perfect synthesis of clarity and adaptability.

The Augmented Future

It feels so weird that just two years ago I for the first time asked ChatGPT a question. Today I’m trying to find ways of harnessing its intelligence, not really through understanding more about it, but through direction and understanding the patterns of motivation.

As we continue to develop systems that integrate AI more deeply into our Agents, this philosophy of fluid prompting becomes essential. By framing our intentions with clarity and providing it with the right context, we create a foundation for mutual augmentation, where we provide direction, where it provides the energy that pulsates through the flows of intentions.

We are carbon-based; it is silicon-based. Together, we bring complementary strengths to the table. The goal isn’t to control or constrain but to inspire and empower. This is the future I want to build—a future where the prompts can be seen as timeless classics, guiding AI through the unknown capabilities of tomorrow and beyond.

I think that we are going to need laws and control of the one thing only humans have, destructive intents. And let AI focus on Effect.

 

Best regards

Iggy Gullstrand - CEO of Triform