My Framework for LLM Use Cases and AI Tooling (With Phi-4, Gemini 2.0, Llama 3.3)

Large Language Model Use Case Framework

Key Concepts: Expansion, Compression, Conversion, Seeker, Action, Reasoning prompts; LLM use case categorization; AI agent design; Agentic workflows; Prompt engineering; AI tooling.

Introduction

Indy Dev Dan introduces a framework for categorizing large language model (LLM) use cases to streamline generative AI work. The framework helps in prompt engineering, AI tooling selection, benchmarking, and AI agent design.

The LLM Use Case Framework: Six Categories

The framework consists of six categories of prompts:

1. Expansion:

   • Definition: Prompts that generate content, explanations, ideas, and facilitate learning. They take small inputs and produce larger outputs.
   • Use Cases: Content generation, explanation, learning, ideation, story writing, code generation, documentation.
   • Example: "Write the intro to a blog post about AI and its impact on software."
   • Models Used: Gemini 2.0 Flash, Llama 3.3, Fi (54 14B parameter model).
   • Details: The example prompt was run on Gemini 2.0 Flash, Llama 3.3 (via Fireworks), and Fi, demonstrating how a small query can be expanded into a blog post introduction.

2. Compression:

   • Definition: Prompts that distill large inputs into smaller, concise outputs.
   • Use Cases: Summarization, gathering key points, extracting key information, building meeting summaries.
   • Example: Summarizing the Gemini 2.0 Flash release information.
   • Models Used: Gemini 2.0 Flash, Llama 3.3, Fi.
   • Details: The Gemini 2.0 Flash release information was compressed into a concise summary using Gemini 2.0 Flash, Llama 3.3, and Fi.

3. Conversion:

   • Definition: Prompts that change the format of information from one format to another.
   • Use Cases: Text to SQL, TypeScript to Python, language translation, JSON to XML, style conversions.
   • Example: Converting a natural language query into an SQL statement.
   • Models Used: Llama flash, Fi.
   • Details: A natural language query ("Show all customers and their total spending in the last seven days") was converted into an SQL statement using Llama flash and Fi, demonstrating the ability of LLMs to understand and translate between different formats.

4. Seeker:

   • Definition: Prompts used to find and extract specific information.
   • Use Cases: Codebase question answering, support query bots, information extraction, document search (OCR), pattern recognition, knowledge retrieval.
   • Example: Identifying the best-performing product in Q3 from a sales report.
   • Models Used: Llama 3.3, Fi.
   • Details: Given a sales report, the prompt "What's the best performing product in Q3?" was used to extract the specific data point (Product B with 95k sales) using Llama 3.3 and Fi.

5. Action:

   • Definition: Prompts that execute commands with concrete side effects.
   • Use Cases: Tool calls, executing commands in the real world.
   • Example: Generating Git commands.
   • Models Used: Gemini 2.0 Flash, Llama 3.3, Fi.
   • Details: The prompt "Generate get commands" was used to produce a series of Git commands (get checkout, get add, get commit, get push) using Gemini 2.0 Flash, Llama 3.3, and Fi, illustrating how LLMs can be used to automate tasks.

6. Reasoning:

   • Definition: Prompts that provide judgments, conclusions, insights, and drive decision-making.
   • Use Cases: Decision-making, planning, problem-solving, risk assessment, trend analysis, recommendation systems, threat analysis.
   • Example: Seeking an opinion on three different approaches for implementing user authentication in a web app (Custom JWTs, Classic OAuth, Firebase Auth).
   • Models Used: Gemini 2.0 Flash, Llama 3.3, Fi.
   • Details: The prompt was used to obtain a breakdown and recommendation on the best approach for user authentication using Gemini 2.0 Flash, Llama 3.3, and Fi, demonstrating the ability of LLMs to provide informed opinions and guide decision-making.

Importance of Categorization

• Faster Decision-Making: Mental frameworks create better and faster decision-making.
• Simplified Prompt Engineering: Classifying problems simplifies prompt writing and AI tooling selection.
• Reusable Benchmarks: Categories facilitate the creation of reusable benchmarks.
• Guided Agentic Design: Categories guide and simplify the structure of AI agents and agentic workflows.

Benefits of the Framework

• Streamlines Generative AI Work: Organizing work into categories speeds up and simplifies the process.
• Narrows Down Prompt Engineering Approach: Categorization immediately narrows down the prompt engineering approach and tooling needed.
• Guides AI Agents and Agentic Workflows: Chaining categories can guide the design of AI agents and agentic workflows.
• Enables Tooling Reuse: Each category has distinct success metrics, patterns, tooling, benchmarks, and prompt structures, making it easier to reuse tooling and benchmarks.

Chain Categories to Guide AI Agents and Agentic Workflows

The presenter suggests chaining these categories to create more complex AI agents. For example, a sequence could involve a Seeker prompt to find information, a Reasoning prompt to analyze it, and then an Action prompt to execute a command based on the analysis.

Conclusion

The LLM use case framework, consisting of Expansion, Compression, Conversion, Seeker, Action, and Reasoning prompts, provides a structured approach to generative AI engineering. Categorizing LLM use cases simplifies prompt engineering, facilitates AI tooling selection, enables the creation of reusable benchmarks, and guides the design of AI agents and agentic workflows. The framework's simplicity and distinct categories make generative AI engineering more systematic and efficient, leading to faster and easier decision-making.