A Taxonomy for AI Prompts

Prompts have become artifacts. When interacting with AI Developer tools, prompts are the cornerstones of our interactions. We can standardize them, or create them on-the-fly. Files like .cursorrules and CLAUDE.md sit alongside source code, shaping how AI tools understand and assist with a codebase. With this proliferation, we need clear ways of classifying them so that it's easier to compare alternatives and generally operate from a higher level.

To help practitioners explore the landscape of prompts for AI Coding, we collected 1,751 prompts from 12 curated sources: GitHub Copilot instructions, Claude configuration files, Cursor rules, and agent system prompts. What emerged was a clear pattern: prompts can be classified along four axes, with Type being the primary discriminator.

Three Types

A prompt's type describes what it fundamentally is. It answers the question: what role does this text play when the AI reads it?

System prompts define behavioural rules and AI identity. These are the most common in our corpus, comprising over 89% of all prompts. They establish who the AI is, set boundaries, and provide context that shapes every subsequent interaction. A System prompt might declare expertise ("You are a senior Svelte developer with expertise in building scalable applications"), set constraints ("Always use TypeScript strict mode"), or establish preferences ("Prefer composition over inheritance").

Task prompts request specific work. They ask the AI to produce something: generate code, fix a bug, explain a concept, or review a pull request. Unlike System prompts that set the stage, Task prompts assign the work.

Meta prompts are prompts about prompting. They define naming conventions for prompt files, contribution guidelines for prompt repositories, or templates for writing effective prompts.

Type	Purpose	Example
System	Behavioral rules defining AI identity	"You are a senior engineer who values clean code"
Task	Immediate work request	"Generate a commit message for these changes"
Meta	Prompts about prompting	Prompt file naming conventions

Prompts in Practice

The simplest prompts to write are behavioural constraints. Hard rules that apply unconditionally:

Always include the language and file name in the info string when you write code blocks.

One line, immediately actionable. This can be part of a System prompt in project files like .cursorrules where you want consistent behaviour across all interactions.

Task prompts can be equally concise when the goal is clear:

Generate a commit message. First, a single sentence under 80 characters. Then after 2 line breaks, up to 5 bullet points, each under 40 characters.

Notice how this combines the request (generate a commit message) with structural requirements (format specification). The type is Task, but it incorporates formatting guidance that shapes the output.

More complex work benefits from explicit structure. A code review prompt might specify not just what to evaluate, but how to communicate findings:

Review the following code, focusing on Readability, Maintainability,
Code Smells, conformance to the Coding Standards.

Tone: Friendly casual tone of a fellow engineer.
Orderly Analysis: Address the code top to bottom.
Provide Examples: For each issue, offer concise insights into potential improvements.

This is a Task prompt with clear expectations about both substance and style. The explicit tone guidance helps capture critical information in a consistent tone and format.

Four Axes

Type is the primary classification, but three additional axes refine how prompts are understood:

Activity describes what work the prompt does. This applies mainly to Task and Skill types:

• Create: Generate or transform content
• Fix: Correct or validate existing content
• Understand: Explain or analyze

Constraint describes how much freedom the AI has:

• Open: AI chooses freely
• Guided: Soft preferences provided
• Bounded: Hard rules with some flexibility
• Scripted: Exact steps to follow

Scope describes how broadly the prompt applies:

• Global: All AI interactions
• Project: This codebase
• Session: This conversation
• Atomic: Single use

Understanding scope helps you decide where a prompt belongs. Constraints that should always apply go in project files. One-off instructions stay in the conversation. Agentic systems codify this.

Explore Further

Browse all 1,751 prompts and filter by type, activity, constraint, and scope:

Prompt Explorer →

The Prompt Explorer includes extensive search-ability and filtering. Not only filter by the Prompt Type and this 4-Axis taxonomy, but also by tech and keyword tags... or just click the random shuffle icon in the detail page to roll-the-dice and find something new.

For detailed definitions of each type and axis:

Taxonomy Reference →

Whether you're writing your first .cursorrules or designing an agent system, understanding these classifications helps you write prompts with intention: knowing not just what to say, but what kind of thing you're saying.