AI Interaction Atlas

A shared language for designing legible, inspectable AI systems.

The AI Interaction Atlas is a structured taxonomy of interaction patterns used to design, reason about, and document human–AI systems. It focuses on how systems behave, where agency lives, and what constraints shape outcomes — not just models or UI.

180 structured elements across 6 system dimensions, designed to make AI systems visible before they are built.

🌐 Explore the Atlas: https://ai-interaction.com
🐙 GitHub: https://github.com/quietloudlab/ai-interaction-atlas

What This Package Is

This npm package provides the Atlas data and utilities:

• A curated, opinionated dataset of AI interaction patterns
• Strongly typed structures for tasks, data, constraints, and touchpoints
• Search, filtering, and validation utilities
• Zero dependencies, tree-shakeable, framework-agnostic

It is designed to be used:

• in code
• in prompts
• in documentation
• in tools built by you or the community

What This Package Is Not

• ❌ Not an AI framework
• ❌ Not a UI kit
• ❌ Not an orchestration engine
• ❌ Not a visual editor

Note:
A visual canvas for composing and mapping systems using the Atlas is being explored separately.
This package intentionally focuses on the language layer, not the tooling.

Who This Is For

You may find the Atlas useful if you are:

• Designing AI-powered products or workflows
• Building internal tools or design systems for AI teams
• Creating AI assistants that reason about interaction patterns
• Documenting or auditing AI systems
• Teaching or learning applied AI system design

You do not need to be an ML engineer to use the Atlas.
It encodes design intent and system behavior, not implementation details.

Why Use the Atlas?

Most AI products are designed at the wrong level of abstraction.

Teams jump to:

“Use an LLM”
“Add an agent”
“Automate this”

Instead of asking:

• Where does human judgment remain essential?
• Which decisions are probabilistic vs deterministic?
• What constraints govern safety, latency, or trust?
• Who is accountable when the system fails?

The Atlas provides a shared vocabulary to answer those questions before systems harden.

Installation

npm install @quietloudlab/ai-interaction-atlas

yarn add @quietloudlab/ai-interaction-atlas

pnpm add @quietloudlab/ai-interaction-atlas

Quick Start

import {
  AI_TASKS,
  searchPatterns,
  getPattern,
  validateWorkflow,
  getAtlasStats
} from '@quietloudlab/ai-interaction-atlas';

// Inspect available AI capabilities
console.log(`${AI_TASKS.length} AI task patterns available`);

// Search patterns by keyword
const reviewPatterns = searchPatterns('review', { dimensions: ['human'] });

// Retrieve a specific pattern
const classifyIntent = getPattern('classify-intent');
console.log(classifyIntent?.description);

// Validate a proposed system workflow
const validation = validateWorkflow([
  'ai_classify_intent',
  'human_review_output',
  'system_log_event'
]);

if (!validation.valid) {
  console.error('Invalid pattern IDs:', validation.invalidIds);
}

// Get Atlas statistics
console.log(getAtlasStats());

The Six System Dimensions

The Atlas models AI systems as compositions of six dimensions:

| Dimension | Description | |---------|-------------| | AI Patterns | Probabilistic capabilities (detect, classify, generate, transform) | | Human Actions | Where human agency lives (review, approve, decide, configure) | | System Operations | Deterministic infrastructure behavior (routing, caching, logging) | | Data Artifacts | Inputs, outputs, and contextual information | | Constraints | Boundaries that shape behavior (latency, privacy, accuracy, cost) | | Touchpoints | Where systems surface (UI, API, voice, notifications) |

Core Exports

Data Collections

import {
  AI_TASKS,
  HUMAN_TASKS,
  SYSTEM_TASKS,
  DATA_ARTIFACTS,
  CONSTRAINTS,
  TOUCHPOINTS,
  LAYERS,
  WORKFLOW_TEMPLATES,
  EXAMPLES,
  ATLAS_DATA
} from '@quietloudlab/ai-interaction-atlas';

Types

import type {
  AiTask,
  HumanTask,
  SystemTask,
  DataArtifactDefinition,
  ConstraintDefinition,
  TouchpointDefinition,
  Layer,
  WorkflowTemplate,
  AtlasData
} from '@quietloudlab/ai-interaction-atlas';

API Overview

searchPatterns(query, options?)

Search across all Atlas elements by keyword.

searchPatterns('review', { dimensions: ['human'], limit: 5 });

getPattern(id)

Retrieve a single pattern by ID (slug or target_id).

getPattern('privacy-compliance');

getPatternsByDimension(dimension)

Retrieve all patterns from one dimension.

getPatternsByDimension('constraints');

validateWorkflow(nodeIds)

Validate that a workflow uses valid Atlas elements.

validateWorkflow([
  'ai_generate_text',
  'human_review_output',
  'system_store_result'
]);

Real-World Uses

• Design audits: Map an existing AI product to surface risks and gaps
• System prompts: Ground AI assistants in real interaction patterns
• Documentation: Generate inspectable system diagrams and specs
• Tooling: Build search, validation, or mapping tools on top of the Atlas
• Education: Teach applied AI system design with concrete language

TypeScript Support

Written in TypeScript with full type definitions.

function analyzeTask(task: AiTask) {
  console.log(task.layer);
  console.log(task.inputs);
  console.log(task.outputs);
}

Bundle & Dependencies

• ~220KB uncompressed
• Tree-shakeable
• Zero runtime dependencies
• Pure data + utilities

License

Apache 2.0 — free to use, modify, and integrate commercially.
See LICENSE for details.

About

Created by Brandon Harwood at quietloudlab —
a design and research studio focused on human-centered AI, system legibility, and responsible interaction design.

• 🌐 https://quietloudlab.com
• 🌐 https://ai-interaction.com

Contributing

The Atlas is open source and evolving.

Contributions, issues, and discussions are welcome — especially around:

• new patterns
• clearer definitions
• real-world examples
• missing constraints

See the GitHub repository for contribution guidelines.