Hi Itamar — congratulations on all these initiatives.
As promised in our private exchange, I wanted to lay out an architectural idea I’ve been exploring for LLM-based applications, which may be useful to others building similar tools. I don’t know how novel this is, but in a world where many tools will increasingly rely on AI, I think it’s a good general practice.
The core idea is simple: all AI prompts live in a dedicated, human-readable folder, separate from application logic.
There are two main reasons for this.
First, radical transparency. If an application makes claims, recommendations, or interpretations that matter ethically or scientifically, then the instructions guiding the AI are part of what should be open to scrutiny. Keeping prompts in a clearly accessible place makes the system legible not only to developers, but also to researchers, ethicists, and communities like EA or academia who may want to understand how conclusions are being generated, not just what the interface shows.
Second, a clean separation between scientific or ethical content and engineering plumbing. Prompts often reflect underlying assumptions, value choices, and ways of thinking about a problem. Keeping them visible and separate from the rest of the code helps ensure that changes to how the AI reasons or frames an issue are intentional and easy to review, rather than happening quietly as a side effect of technical work. In practice, this folder is meant to be the main reference for what the AI is told to do, while the surrounding code simply handles running it.
In our Welfare Food Explorer app, for example, this structure allows researchers and non-developers to easily find, read, and reason about what the AI is being instructed to do, without needing to navigate the rest of the codebase.
We adopted this approach because in applications that touch science, ethics, welfare, or normative interpretation, how the AI reasons is part of the substance of the system itself. Making prompts visible, inspectable, and discussable helps treat AI behavior as something that can be examined, debated, and improved by a broader audience.
I hope this perspective is useful to others. Cheers.
Hi Itamar — congratulations on all these initiatives.
As promised in our private exchange, I wanted to lay out an architectural idea I’ve been exploring for LLM-based applications, which may be useful to others building similar tools. I don’t know how novel this is, but in a world where many tools will increasingly rely on AI, I think it’s a good general practice.
The core idea is simple: all AI prompts live in a dedicated, human-readable folder, separate from application logic.
There are two main reasons for this.
First, radical transparency. If an application makes claims, recommendations, or interpretations that matter ethically or scientifically, then the instructions guiding the AI are part of what should be open to scrutiny. Keeping prompts in a clearly accessible place makes the system legible not only to developers, but also to researchers, ethicists, and communities like EA or academia who may want to understand how conclusions are being generated, not just what the interface shows.
Second, a clean separation between scientific or ethical content and engineering plumbing. Prompts often reflect underlying assumptions, value choices, and ways of thinking about a problem. Keeping them visible and separate from the rest of the code helps ensure that changes to how the AI reasons or frames an issue are intentional and easy to review, rather than happening quietly as a side effect of technical work. In practice, this folder is meant to be the main reference for what the AI is told to do, while the surrounding code simply handles running it.
In our Welfare Food Explorer app, for example, this structure allows researchers and non-developers to easily find, read, and reason about what the AI is being instructed to do, without needing to navigate the rest of the codebase.
We adopted this approach because in applications that touch science, ethics, welfare, or normative interpretation, how the AI reasons is part of the substance of the system itself. Making prompts visible, inspectable, and discussable helps treat AI behavior as something that can be examined, debated, and improved by a broader audience.
I hope this perspective is useful to others. Cheers.