title: "Architecting Bespoke Productivity Ecosystems: Leveraging LLM-Driven Simulations for UX Optimization" date: 2026-10-20 tags: [ai, software_development, ux_design, productivity_engineering] description: "An exploration of using Large Language Models to audit complex workflows and develop custom, keyboard-first applications."
Architecting Bespoke Productivity Ecosystems: Leveraging LLM-Driven Simulations for UX Optimization
In the modern era of digital productivity, developers and power users face a persistent dilemma: do you attempt to force-fit your unique cognitive workflows into existing SaaS ecosystems like Notion, Tana, or Todoist, or do you undertake the technical debt of building a custom application? While the latter is often dismissed due to the perceived complexity of software maintenance, the emergence of advanced Large Language Models (LLMs) like Claude has fundamentally shifted the cost-benefit analysis.
This post explores a methodology for using AI not just as a coding assistant, but as an automated QA engineer and UX auditor capable of simulating end-to-end user journeys to identify and eliminate workflow friction.
The Friction Audit: Using LLMs as UX Researchers
The primary challenge in any complex productivity system—whether it is a combination of disparate tools or a unified self-built application—is the accumulation of "noise." In my own development of an all-in-one productivity suite, early iterations (v0.1) suffered from excessive field density and cognitive overload.
The breakthrough came when I transitioned from manual debugging to an AI-driven audit. By providing an LLM with a detailed breakdown of the current toolstack—including the integration points between task management, calendars, planners, and high-volume databases (such as health data containing millions of entries)—I was able to prompt the model to identify specific friction points.
The objective is to move beyond simple feature requests and toward "frictionless" interaction. This involves analyzing:
- Information Retrieval Latency: How many steps/keystro's are required to find a specific node?
- Input Fragmentation: Are there disconnected silos between the inbox and the permanent knowledge base?
- UI Noise: Which UI elements are visible but non-essential for daily operations?
The ICO Framework: A Blueprint for System Architecture
To provide the AI with enough context to perform a meaningful audit, one must utilize a structured framework. I employ the ICO (Integrated Cognitive Organization) Framework, which defines productivity through four distinct quadrants:
- Personal Knowledge Management (PKM)
- Personal Project Management (PPM)
- Business Knowledge Management (BKM)
- Business Project Management (BPPM)
By mapping tools and workflows to these four circles, you create a semantic map that can be fed into an LLM. This allows the AI to understand not just what tools are being used, but the intent behind the data flow. When the AI understands the relationship between your "Inbox" (asynchronous processing) and your "Thought Lab," it can suggest optimizations like unified Command K interfaces or streamlined node creation via Cmd+N.
Advanced Simulation: Persona-Based Stress Testing
Perhaps the most transformative application of LLMs in this workflow is the ability to run Persona-Based Simulations. Rather than manually testing every new feature, I generate a synthetic persona—for example, "Scott Bennett"—complete with a mock dataset including CRM entries, project timelines, and daily routines.
I then instruct the AI to simulate an entire day in the life of this persona within the application. The LLM acts as an agent navigating the software's logic, attempting to execute specific tasks:
- "Find the Quality of Earnings report." (Result: Failure due to disconnected database links between files and mail).
- able to "Capture a quick note during a meeting interruption." (Result: Friction identified because there was no rapid-capture mechanism).
This simulation identifies "broken" workflows—such as a habit tracker that can only be updated on its own specific page, requiring unnecessary navigation—before a single line of production code is even deployed. This allows for the identification of high-latency paths, such as excessive Shift+Tab usage, which can then be optimized via keyboard-first shortcuts (e.g., implementing an F6 keybind).
From Micro-Utilities to Full-Scale Applications
The development process follows a modular trajectory. It begins with "Utility Apps"—small, functional scripts designed to solve immediate problems. An example is my "Agent Watch" utility: a lightweight dashboard that monitors running background servers and agents within various Claude Code sessions. This was built in approximately 10 minutes by prompting Claude to create a functional interface for killing/managing processes.
As these utilities prove their value, they can be integrated into a larger, unified architecture. The goal is an application characterized by:
- Keyboard-First Design: Minimizing mouse dependency through robust Command K systems and streamlined navigation.
- Contextual UI (Progressive Disclosure): Using "More" sections to hide deep metadata, keeping the primary interface clean while maintaining data availability.
- Integrated Data Silos: A unified CRM that connects directly to email streams, paired with a note-taking system that utilizes
@mentionsfor instant entity linking.
The Compounding Effect of AI-Generated Documentation
The final stage of this ecosystem is the automation of the learning loop. Once the interface and code are established, the LLM can use that very same code to generate animated tutorials. By feeding the UI logic back into Claude and utilizing 11 Labs for high-quality voiceovers, I can create personalized, automated training modules. This ensures that as the application evolves (e.g., moving from v2.01 to v2.4), the documentation scales alongside the codebase without manual intervention.
Conclusion: The Shift in Software Ownership
The barrier to entry for bespoke software development has collapsed. We are no longer limited by our ability to write syntax, but by our ability to define requirements and audit workflows. While established tools like Todoist remain excellent for general use, the power of building your own lies in the ability to eliminate the specific friction points that generic software cannot address. By leveraging LLM-driven simulations, you can build a highly polished, personalized productivity engine in a fraction of the time previously required.