Use Case Taxonomy

A systematic classification of generative AI patterns for software development and content creation.

Overview

• 📦 Basic Patterns — Foundational generation patterns: creation, transformation, and synthesis
• 🔄 Advanced Patterns — Incremental, hierarchical, and iterative generation strategies
• 🔧 Combined Workflows — Real-world pipelines combining multiple patterns

Basic Generative Use Cases

0 → 1: Content Creation (Foundational)

Generate new content from scratch based on descriptions or prompts.

Description → LLM → Generated File
"Create a REST API for user management" → LLM → UserController.kt

Examples:

• Generate file from description
• Generate images from text prompts
• Create boilerplate code from specifications
• Generate documentation from requirements

1 → 1: Content Transformation (Transform)

Transform existing content while preserving its essential meaning or structure.

Input → Transform → Output

Examples:

• Summarization of articles
• Translation between languages
• Edit / feature addition to existing code files
• Code refactoring and optimization

(A + B + …) → C: Synthesis (Combine)

Combine multiple inputs to create unified, coherent output.

[A, B, C] → Synthesize → Result

Examples:

• Synthesis by Example: Generate new content based on multiple input examples
• Pattern Extraction: Identify common patterns from multiple inputs
• Mass Summarization: Summarize multiple documents into a single cohesive summary

(A − B + C) → D: Analogical Generation (Advanced)

Create new content by combining and modifying existing pieces through analogy.

A − B + C → D

Concept: Transfer features from one class/file to another — a sort of "transgenic" content generation.

UserService (A) - BasicCRUD (B) + CachingLayer (C) → CachedUserService (D)

Advanced Generative Use Cases

A → A + δ: Incremental Generation (Extend)

Generate additions to existing content, continuing from where it left off.

Existing (A) → Continue → Extended (A + δ)

Key Insight: Essentially telling the model to "continue" from where it left off, useful for long-form content generation.

Hierarchical Generation (Scale)

Generate complex content through multiple stages of refinement and transformation.

Spec
├── Module A
│    ├── File 1
│    └── File 2
├── Module B
│    ├── File 3
│    └── File 4
└── Module C
├── File 5
└── File 6

Key Insight: Allows for exponential growth in content volume. LLM output is generally constrained by input size, but hierarchical generation bypasses this limitation.

Applications:

• Spec-driven development
• Large codebase generation
• Documentation hierarchies

Iterative Editing (Refine)

Repeatedly refine content through multiple editing passes.

A₀ → Edit → A₁ → Edit → A₂ → ... → Aₙ

Considerations:

• Patching Strategy: Full replacement vs. modification of existing content
• Convergence: Monitor for stability and diminishing returns
• Quality Gates: Define stopping criteria

Error Correction (Fix)

Identify and correct errors using a feedback loop with external validation.

Code → Validator (Tests / Linter)
❌ Fail → Fix → (loop back to Validator)
✅ Pass → Done

Considerations:

• May require many iterations to achieve desired quality
• Periodic human interventions may be necessary
• Set maximum iteration limits to prevent infinite loops

Combined Workflows

Real-world workflows often combine multiple patterns to achieve complex goals.

Example: Spec-Driven Development Pipeline

Spec → (Modules)            [Hierarchical]
↓
Each Module → Code          [0→1 Creation]
↓
Code → Tests                [1→1 Transformation]
↓
while(tests fail) → Fix     [Error Correction]

Quick Reference