Fuzzy Gates

Fuzzy Gates employs a prompt style known as Transitive Task Trees (T3).

The initiating prompt includes an initial query context—a robust guiding statement, describing the motivation, audience, constraints, etc, and transitive seed instructions (e.g., Make a double cheeseburger). The AI then generates a detailed task tree, breaking down the seed instructions into a series of transitive subtasks during recursive evaluation.

Quickstart

The dashboard Web client will initialize on port 3010 once the cloned repo is deployed and running in Docker. You must include a .env file. The OpenTelemetry API key is optional, but the OpenAI API Key must be provided.

Requirements

• Docker (Desktop 4.24/Docker Compose 2.22.0)
• Node

Get started

• npm run docker:up - Build
• npm run open:redis - Open RedisInsight [password is key_admin]
• npm run docker:reset-redis - Reset Redis [reset database]

Once the container builds and is actively running, navigate to the Web App at http://localhost:3010/tasks. RedisInsight is included for browsing the raw data.

Principles

T3 is based on the premise that transitive verbs can elicit process intelligence from an LLM. While 'be' verbs are linked with terminal narrative states, transitive verbs propel the story forward.

BE Verbs describe terminal states of being or becoming and include verbs like: be, am, is, are, was, were, been, have, has, had, do, does, did, can, could, may, might, shall, should, will, would, must.

Strategy

T3 is a multi-phase process that uses an evolving context window and a set of gestalt-oriented strategies so the LLM can automatically refine the task tree; however, it also supports direct user influence at any juncture in the task tree, allowing users to influence the parse details, depth, and ultimate exection.

• I. GENERATE the Transitive Task Tree. Recursively clarify or subtask the instructional seed using an evolving context:
  • INITIAL CONTEXT
  • ANCESTOR TASK LIST
  • PREVIOUS SIBLING TASK
  • FOLLOWING SIBLING TASK
• II. ITERATIVELY IMPROVE the Transitive Task Tree. Use analytics to generate a comprehensive view of the task set to:
  1. PRUNE TASKS
  2. CLARIFY TASKS
  3. EXPAND (SUBTASK) TASKS
• III. Execute the Transitive Task Tree. Choose a target model appropriate to generate the output (like choosing a model trained on code to author modules.)

Example 1: Recipe for Classic Chicken Pot Pie

INPUT/PROMPT

• context All units are American. Specify exact quantities and recommend specific varietals that reflect a NorCal/Napa style palette (fresh); recommend Kosher salt, for example, never salt. BE concise! Never overexplain! Explain how to prep and cook, but never more! Recipe book style! Limit to 4 levels of task breakdown. Assume they are already present and gathered.
• instructions Make classic chicken pot pie with peas carrots and corn with extra flaky crust.

One Shot Prompt OUTPUT

One Shot generates the following output using OpenAI's GPT-4o model, providing a single, comprehensive response that includes all necessary steps along with ingredients to complete the task. output

OpenAI-o1-preview OUTPUT

o1-preview generates the following output using OpenAI's OpenAI-o1-preview model, providing a single, comprehensive response that includes all necessary steps along with ingredients to complete the task. output.

T3 Prompt OUTPUT

T3 produced significantly more detailed outpu using the same GPT-4o model, particularly at the task level with detailed inputs and outupts described for every step. output

Example 2: GitHub Pages Site with React and AWS Route 53

INPUT/PROMPT

• context The project is focused on creating a GitHub Pages site using React, targeted at personas including CIO, VP of Engineering, Engineering Manager, Principal Architect, DevOps Engineer, and Cloud Ops Engineer. The value proposition (content for the site) centers on reducing the need for experienced engineering talent. The customer has already purchased the domain from AWS (Route 53) and requires it to be mapped to the GitHub Pages site. Break down the project into major tasks like development, setup, deployment, and connecting, with a focus on simplicity and clarity. Assume that all necessary tools and accounts are already available.
• instructions Develop a static site using React, structured around key personas in the tech industry. Set up the GitHub repository for GitHub Pages, ensuring it's ready for web hosting. Deploy the site, ensuring all React components and static assets are properly configured for GitHub Pages. Connect the custom domain purchased from AWS Route 53 to the GitHub Pages site, including DNS configuration and verification of domain ownership. Return all instructions in Markdown as a single response users can follow step-by-step.

One Shot Prompt OUTPUT

One Shot generates the following output using OpenAI's GPT-4o model, providing a single, comprehensive response that includes all necessary steps to complete the task. output

T3 Prompt OUTPUT

T3's iterative back-and-forth produces significantly more detailed output, resulting in a richer and more nuanced task breakdown. output

Conclusion

While one-shot strategies can be compared to throwing a perfectly chiseled chunk of marble into a lake and observing the emerging ripples, T3 is more akin to skipping a stone across the water, letting the water fragment the stone along its journey. This approach is interactive, dynamic, and iterative. The outcome, while not always perfect, is significantly more detailed and extensible.

Run

Use the POSTMAN files (included in the /docs folder) to call the APIs.