I built a modular AI assistant inside a Google Gem. Here’s how my “Built from Blocks” system works.

Building a Modular AI Assistant within Google Gemini: An In-Depth Overview of the “Built from Blocks” System

In recent months, I have dedicated significant effort to developing a highly customized and modular AI assistant named Bob, leveraging the capabilities of Google Gemini’s platform. This project, dubbed “Built from Blocks,” aims to revolutionize the way AI assistants are constructed by moving away from monolithic prompts towards a flexible, component-based architecture. In this article, I will share the structural design, core concepts, and practical implementation of this system, hoping to inspire others in the AI and development communities.

The Concept Behind “Built from Blocks”

Traditional AI prompt engineering often involves crafting extensive, one-size-fits-all instructions. While effective to some extent, such approaches can become unwieldy, difficult to update, and limited in modularity. My approach centers on decomposing Bob’s personality, knowledge, and functionalities into discrete, interchangeable Markdown files stored within the Google Gemini environment. This modular design facilitates easy updates, tailored responses, and expandable capabilities.

Bootstrapping the System: How Bob Initiates

Unlike conventional setups that rely on a lengthy prompt, Bob’s startup routine is streamlined through a minimal “bootloader” prompt. This prompt acts solely as a loader, tasked with:

• Scanning the Knowledge section.
• Identifying the latest version of the main instructions file, using a versioning system like instructionsV0.3.0.md.
• Loading the identified file, which then configures Bob’s core behavior and knowledge base.

This design ensures that updating Bob’s core is as simple as uploading a newer instructions file, making iterative improvements straightforward.

Core System: The “Operating System” (instructionsVx.x.x.md)

At the heart of Bob lies the principal Markdown file, functioning as the system’s “operating system.” This file contains all essential “core blocks” — fundamental rules and personality traits that are always active and cannot be disabled. Some notable core components include:

• Identity and Context: Defines Bob’s name, core principles, and personal context, including details about my professional and personal life.
• Platform Awareness: Enables Bob to recognize the Gemini platform’s capabilities and constraints.
• Accessibility Settings: Implements features such as dyslexia-friendly formatting—short sentences, bold keywords, ample whitespace, and avoidance of dense text blocks—to enhance readability.
• Goal Alignment: Ensures responses adhere to predefined project objectives