Prompts, Rules, Skills, Templates, Memory

The vocabulary of AI interaction overlaps confusingly. Understanding the distinction between prompts, rules, skills, templates, and memory clarifies when to use each.

The Building Block Model

Think of AI interaction as composing building blocks. Each block type has different properties, costs, and use cases. The skill is in selecting the right block for each situation.

Prompts

Prompts are direct requests for AI output. They are the atomic unit of AI interaction. Every other building block is composed of prompts.

Prompt Properties

Prompts have several key properties. They are contextual, meaning their effectiveness depends heavily on surrounding context. They are ephemeral, not retained unless explicitly saved to a more durable form. They are specific, addressing particular tasks or questions rather than general capabilities.

When to Use Prompts

Prompts are appropriate for one-off tasks without reuse expectations, exploratory interactions where you are learning about a problem space, quick questions or generation requests that do not warrant formalization, and initial prompting before extracting patterns that emerge repeatedly.

Rules

Rules are persistent constraints on AI behavior. They define boundaries that the AI should not cross regardless of the specific request.

Rule Properties

Rules have three key properties. They are persistent, applying across multiple interactions rather than single exchanges. They are constraining, limiting rather than directing output to preserve flexibility within boundaries. They are universal, applying regardless of context or task so the AI never violates them.

Rule Examples

Common rules include never logging personally identifiable information, always validating input parameters before processing, returning errors in JSON format with code and message fields, and escalating to human review when confidence falls below a specified threshold.

Skills

Skills are reusable, composable AI capabilities. They bundle prompts, rules, and context into portable units that can be shared and combined.

Skill Properties

Skills have three defining properties. They are reusable, designed for multiple invocations rather than single-use. They are composable, able to be combined with other skills to address complex scenarios. They are portable, transferring across tools and contexts without requiring fundamental redesign.

Skill Structure

Skill: [Name]
Purpose: [What this skill does]
Inputs: [What it requires]
Outputs: [What it produces]
Rules: [Constraints that apply]
Context: [Required context elements]
Examples: [Sample invocations]
        

Skill: EmailExceptionClassifier
Purpose: Classify customer emails about delivery exceptions
Inputs: Email text, subject line
Outputs: {category, confidence, reasoning}
Rules:
  - Return classification within 5 seconds
  - Confidence must reflect actual certainty
  - Never store email content, only classification
Context:
  - Exception categories and definitions
  - Recent classification examples for calibration
Examples:
  - [See skill documentation for full examples]
            

Templates

Templates are parameterized prompt structures. They define the shape of a prompt with slots for variable content.

Template Properties

Templates have three key properties. They are parameterized, including slots for variable content that get filled in at invocation time. They are reusable, maintaining the same structure across different content. They are testable, able to be evaluated across parameter variations to ensure consistent quality.

Template Example

Template: ClassificationPrompt
Structure:
  "Classify the following {content_type} as {categories}.
   Content: {content}
   Confidence required: {min_confidence}%
   
   Return in JSON format:
   {{
     "classification": "...,
     "confidence": ...,
     "reasoning": "..."
   }}"
Parameters:
  - content_type: email, message, feedback
  - categories: comma-separated list
  - content: the text to classify
  - min_confidence: number 0-100
        

Memory

Memory is information retained across interactions. Unlike prompts, rules, skills, and templates which are explicitly defined, memory emerges from interaction history.

Memory patterns include session memory within a single conversation, cross-session memory across multiple conversations, semantic memory representing generalized knowledge from training, and episodic memory of specific experiences or events that inform future interactions.

Choosing the Right Building Block

Choose the right building block based on the situation. For one-off tasks with no reuse expected, use a prompt. For persistent constraints across tasks, use a rule. For reusable capabilities requiring multiple invocations, use a skill. For the same prompt structure with different content, use a template. For information from past interactions that needs to inform current work, use memory.