AI & design

1. What Is Artificial Intelligence?

Artificial intelligence is not a single technology. It is a broad field of computer science focused on building machines that can perform tasks typically requiring human intelligence — understanding language, recognizing images, making decisions, and generating creative content.

To understand where the tools you use fit, it helps to picture AI as a set of nested circles:

Artificial Intelligence is the outermost circle — the entire field.

• Inside it sits Machine Learning (ML) — a subfield where systems learn patterns from data rather than following explicit programming.
• Inside ML sits Deep Learning — a subset that uses neural networks with many layers to process complex data like images, audio, and text.
• Inside Deep Learning sit Foundation Models — large-scale neural networks pretrained on enormous datasets that can be adapted to many different tasks.
• Inside Foundation Models sit Large Language Models (LLMs) — text-specialized systems like GPT-5.4, Claude 4.6, and LLaMA 4 that power the chatbots and writing assistants you use every day.

Narrow AI vs. General AI. Every AI tool available today is a form of narrow AI — designed for specific tasks. A model that generates images cannot write code. A model that transcribes audio cannot design a website. General AI — systems with human-like flexibility across any domain — remains theoretical and is not a practical concern for creative professionals in 2026.

Why this matters. When a vendor claims their product uses "AI," understanding which layer of this stack they operate at tells you something about what the tool can and cannot do. A "deep learning" image enhancer operates at a different level of capability than a "foundation model" text generator.

2. Generative AI vs. Discriminative AI

Not all AI systems create content. Some classify it. The distinction between generative and discriminative AI determines whether a tool produces something new or evaluates something existing.

Generative AI learns the full probability distribution of its training data — the complete landscape of what is possible. Once it understands that landscape, it can sample new outputs that resemble the training set but have never existed before. Midjourney generating an image, Suno composing a song, and GPT-5.4 drafting a strategy document are all generative processes. These models are creators.

Discriminative AI learns only the boundaries between categories — the lines that separate one thing from another. An email spam filter, a photo face-detection system, and a sentiment analyzer that determines whether a tweet is positive or negative are all discriminative. These models are evaluators. They do not create new content. They classify existing content.

A useful analogy. Imagine two artists in a museum. The first studies every painting in exhaustive detail — brushwork, color palette, composition, subject matter. After years of study, they return to their studio and paint something entirely new that embodies everything they learned. That is generative AI. The second artist becomes an expert at spotting forgeries. They learn the precise differences between authentic works and fakes. They can tell you definitively whether a painting is genuine, but they cannot paint one themselves. That is discriminative AI.

Generative Adversarial Networks (GANs) combine both approaches. A GAN consists of two competing neural networks — a generator that creates synthetic outputs and a discriminator that judges whether those outputs are real or fake. The generator tries to fool the discriminator; the discriminator tries to catch the generator. This competition drives both to improve continuously. GANs were the dominant generative approach before diffusion models emerged and are still used in specialized applications.

3. Large Language Models vs. Foundation Models

This is the most commonly confused distinction in contemporary AI discourse. The terms are used interchangeably in marketing and media, but they describe different levels of a hierarchy.

A foundation model is a large-scale neural network pretrained on broad data that can serve as the base for many downstream tasks across potentially multiple data modalities — text, images, audio, video, or code. Think of it as a general-purpose engine.

A large language model (LLM) is a text-specialized foundation model. It is one type of foundation model, focused specifically on language understanding and generation.

The relationship is hierarchical: all LLMs are foundation models, but not all foundation models are LLMs.

A helpful metaphor is a tree. The foundation model is the thick trunk, supporting many branches:

• Text-only branches (LLMs): GPT-5.4, LLaMA 4, Mistral Large 3 — models that read and write text.
• Multimodal branches: GPT-5.4, Gemini 3.1 Pro, Claude Opus 4.6 — models that process and generate text, images, audio, and video within a single architecture.
• Vision-only branches: CLIP (image understanding), SAM (image segmentation) — models focused purely on visual tasks.
• Audio branches: ElevenLabs, Whisper — models specialized in speech and sound.

As of early 2026, the line between LLM and multimodal foundation model has blurred significantly. GPT-5.4 and Claude 4.6 are natively multimodal — they handle text, images, audio, and video interchangeably within a single conversation. Calling them "language models" understates their capability, but the term persists because language remains their primary interface.

Why this matters for your work. When a client asks whether to invest in an "LLM solution," understanding this distinction helps you ask better questions. Do they need text generation? Image analysis? Video understanding? Audio processing? The answer determines which branch of the foundation model tree to explore — and which tools to recommend.

4. AI Workflows vs. AI Agents

This distinction increasingly determines enterprise AI procurement, project architecture, and governance requirements. Understanding it is essential for advising clients and structuring your own AI-assisted projects.

An AI workflow is a predefined, deterministic sequence of steps orchestrated by explicit code logic. The path is fixed. If step A produces output X, step B always follows. Workflows are predictable, auditable, and reliable for well-understood processes. A workflow that extracts text from a PDF, summarizes it, and emails the summary to a distribution list follows the same path every time. There is no decision-making — only execution.

An AI agent is a system where an LLM dynamically directs its own process. Given a goal, the agent selects tools, makes decisions, and adapts its approach based on intermediate results. An agent tasked with "research this company's competitive landscape" might decide to search the web, read annual reports, analyze social media sentiment, and synthesize findings — all autonomously, with the LLM deciding at each step what to do next.

A side-by-side comparison:

The hybrid reality. Most production AI systems in 2026 use a hybrid architecture — workflows provide structural reliability for known processes, while agents handle adaptive decision-making within contained steps. A customer service system might use a workflow to authenticate callers and retrieve account data (predictable steps), then hand off to an agent to handle the actual problem-solving (adaptive conversation). This gives you the reliability of workflows where you need it and the flexibility of agents where it matters.

A practical analogy. A workflow is a conveyor belt in a factory — packages move through fixed stations in a fixed order: scan, label, sort, dispatch. Every package follows the same path. An agent is a courier given a destination and a deadline — they choose the route, the mode of transport, and the intermediate stops based on real-time traffic, weather, and their own experience. Both deliver packages. The conveyor belt is faster and more reliable for standard items. The courier is essential for anything that does not fit the standard process.

5. Generative AI, Agentic AI & AI Agents

Three paradigms represent a spectrum of increasing autonomy in AI systems. Understanding where a tool falls on this spectrum determines how you should use it, what governance it requires, and how you communicate about it to clients.

Generative AI — The Tool

This is the entry point. A generative AI system responds to a single prompt, produces output, and stops. There is no memory between sessions unless explicitly engineered. There is no ability to check external facts, browse the internet, or take actions in other software. The model generates and halts.

• Examples: Midjourney producing an image, GPT-5.4 writing a paragraph, Suno generating a song
• Interaction pattern: Human gives instruction → AI produces output → interaction ends
• Governance needed: Quality review of outputs; fact-checking; IP clearance
• Creative role: Execution assistant — you direct, it produces

Agentic AI — The Connected Assistant

Agentic AI introduces the ability to integrate with external tools, APIs, and data sources. The user still specifies the task, but the system can choose which model to use, which APIs to call, and how to sequence operations. It operates within predefined logic and system guidelines — it executes tasks that have been programmed, but does not independently set goals.

• Examples: A system that checks your calendar, finds available slots, and books a meeting; a tool that retrieves documents from a database and summarizes them
• Interaction pattern: Human gives task → AI selects tools and executes within guidelines → reports results
• Governance needed: Tool access permissions, API key management, output review
• Creative role: Connected assistant — it has more capabilities but still follows your playbook

AI Agents — The Autonomous Partner

AI agents specify their own tasks (or interpret high-level human goals into specific tasks), fetch data from multiple sources, design multi-step processes autonomously, adapt based on feedback, and produce verified results that may trigger further actions. They operate through a continuous OODA loop:

• Observe: Gather information from the environment — user input, data feeds, sensor data, real-time changes

• Orient: Apply learning from past interactions, recognize patterns, contextualize new information

• Decide: Evaluate options, weigh goals against constraints, plan actions, assess risk

• Act: Execute decisions, monitor effects, capture feedback to inform the next cycle

• Examples: An autonomous research agent that spends an hour gathering and synthesizing competitive intelligence; a system that monitors brand mentions, detects emerging crises, and drafts response recommendations

• Interaction pattern: Human gives goal → AI autonomously plans and executes → reports with reasoning

• Governance needed: Comprehensive oversight frameworks, human approval gates, audit logging, clear escalation paths

• Creative role: Strategic partner — it can operate independently but requires human judgment for consequential decisions

Decision Framework

When evaluating an AI tool for a project, ask:

1. Is this purely generative (one-shot output)? → Quality review is sufficient
2. Is this agentic (multi-step but predefined)? → Test the workflow, verify integrations
3. Is this a full agent (autonomous goal pursuit)? → Establish governance framework, set approval gates

The answer determines your governance requirements, human oversight needs, and client communication strategy.

6. Prompt Engineering

Prompt engineering is not typing questions into a chatbot. It is a structured communication discipline that determines whether an AI tool produces generic output or transformative creative material. Understanding how prompts work gives you precise control over AI output.

How Prompts Work

When you submit a prompt, it is tokenized — broken into sub-word units that the model processes mathematically. The model's attention mechanism distributes its computational focus across these tokens according to learned patterns. Words and phrases near the beginning of a prompt typically receive higher attention weight than those at the end. Structural markers (headings, bullet points, code fences) create attention boundaries that the model uses to organize its reasoning.

This is not metaphorical. It is literally how transformer architectures function. Understanding this explains why prompt structure matters as much as prompt content.

Core Techniques

Zero-Shot Prompting

Provide a task without examples. Effective when the task is unambiguous and within the model's training distribution.

Example:

Analyze the brand positioning of [Company X] based on their website copy. 
Identify the primary value proposition, target audience signals, and emotional territory. 
Format as a table with columns: Element, Observation, Strategic Implication.

This works because the task is clear, the context is specific, and the output format is precisely defined.

Few-Shot Exemplar Prompting

Provide two to five examples of the desired input-output pattern before asking the model to generate a new example. This dramatically improves output quality for stylistic or format-specific tasks.

Example:

Generate brand tagline variants in the following style.

Example 1
Brand: Patagonia
Tagline: "We're in business to save our home planet."

Example 2
Brand: Nike
Tagline: "Just Do It."

Example 3
Brand: [Client Name]
Tagline:

The model recognizes the pattern — concise, brand-aligned, emotionally resonant — and generates accordingly. The quality of examples directly determines the quality of output.

Chain-of-Thought (CoT) Prompting

For complex analytical tasks, explicitly request that the model reason step by step before providing its final answer. This forces the model to allocate reasoning capacity to intermediate steps.

Example:

When evaluating market entry strategy for [Client], first identify the three most 
relevant market segments. For each segment, analyze barriers to entry, competitive 
intensity, and brand differentiation potential. Then synthesize a recommended 
prioritization with justification. Think step by step.

Role-Based Prompting

Assign the model a specific persona to activate relevant knowledge and tone patterns.

Example:

You are a senior brand strategist with 20 years of experience in luxury goods. 
You specialize in cultural trend analysis and have advised LVMH, Hermes, and Rolex. 
Review the following consumer research findings and provide strategic recommendations 
through the lens of luxury brand building.

Use this technique responsibly. The role is a prompt engineering technique for output quality, not a claim of actual expertise.

Negative Prompting (Image Generation)

In image generation tools like Stable Diffusion and Midjourney, negative prompts specify what should not appear in the output.

Example:

Professional product photography, luxury cosmetics, soft studio lighting, 
clean white background --no blurry, distorted, ugly, deformed hands, 
watermark, text, logo

This eliminates common generation artifacts that would otherwise require hours of retouching.

Prompt Engineering Best Practices

1. Lead with the task. Put your core request at the beginning of the prompt. Background context should follow, not precede, the core task.
2. Specify output format. Tell the model exactly how you want the response structured — table, bullet list, numbered steps, JSON, etc.
3. Set constraints. Define what to include and what to exclude. Constraints improve precision.
4. Use delimiters. Separate instructions, examples, and source material with clear markers (XML tags, code fences, headers).
5. Iterate. The first prompt rarely produces the best output. Refine based on what you receive.

7. Markdown — The Language of AI

Markdown has emerged as the universal format for human-AI collaboration. Every major AI system — ChatGPT, Claude, Gemini, Cursor, Perplexity, GitHub Copilot — processes Markdown natively. Understanding Markdown is a fundamental skill for anyone working with AI.

Why AI Systems Prefer Markdown

Large language models are trained on enormous corpora of text, and Markdown is disproportionately represented in high-quality training data — technical documentation, README files, academic papers, wiki pages, and structured discussions. When you write prompts in Markdown, you speak the model's native structural language. A prompt formatted with clear headings, bullet lists, and code blocks receives better attention-weighting than unstructured text.

Essential Markdown Syntax

Best Practices for Markdown in AI Prompts

1. Structure before content. Outline your prompt using Markdown headings before filling in details. A prompt with clear # Task, ## Context, and ## Output Requirements sections consistently outperforms unstructured text.
2. Use code fences for examples. Wrap example outputs in triple backticks. This creates a clear boundary between instructions and examples.
3. Tables for structured comparison. When requesting comparative analysis, provide or request Markdown tables. Tables force structure and eliminate ambiguity.
4. Blockquotes for source material. When feeding client feedback or research findings into a prompt, use the > blockquote syntax. This separates source material from your instructions.
5. Headers as hierarchy. Use heading levels to establish importance. Reserve # for the core task, ## for context and constraints, ### for examples and edge cases.
6. Save effective prompts as .md files. Build a personal library of prompt templates organized by task type. A well-crafted prompt is a reusable asset.

8. Context Windows, Tokens & Working Memory

Every AI model has a finite context window — the maximum amount of text it can process in a single conversation or prompt. Understanding context windows is essential because it determines how much reference material, how many examples, and how much conversation history a model can retain.

What Is a Token?

A token is the atomic unit of text processing for language models. It is not a word. As a rough guide, one token equals approximately 0.75 words in English, or about four characters. "Token" itself is one token. "Internationalization" might be split into three tokens. Understanding tokenization matters because it explains why certain words are processed differently — common words are often single tokens; rare technical terms may be split into multiple subword tokens.

Current Model Context Windows (2026)

Best Practices for Context Management

1. Lead with what matters. Because attention mechanisms weight early tokens more heavily, place critical instructions at the beginning. This matters less with newer models (GPT-5.4, Claude 4.6) but still improves reliability.
2. Summarize instead of appending. When a conversation grows long, summarize key decisions rather than maintaining the full transcript. This preserves essential context while freeing token capacity.
3. Chunk large documents. For documents exceeding the context window, break them into thematically coherent sections and process each with a consistent framework, then synthesize results. Less critical with Llama 4 Scout's 10M window, but valuable for cost control.
4. Use structured reference. Restate critical constraints in each prompt rather than assuming the model remembers from fifty messages ago.
5. Monitor token usage. API costs vary significantly. GPT-5.4 and Claude Opus 4.6 are premium-priced; Gemini 3.1 Pro and DeepSeek V3.2 offer more cost-efficient alternatives for high-volume work.

9. Understanding AI Output Formats

Creative professionals must understand not just what AI tools produce, but in what format they produce it. A Midjourney image and a Stable Diffusion image may look similar, but one arrives as a compressed JPEG while the other can output a layered TIFF or depth map.

Raster Image Formats

AI image generators typically output JPEG (compressed, lossy, web-ready), PNG (lossless with transparency support), or WebP (modern compressed format). For production design work, always request PNG or source files. JPEG compression artifacts become visible at print scale and during color correction.

Vector Formats

Tools like Vectorizer.AI, Adobe Illustrator's vectorize feature, and Recraft output SVG (scalable vector graphics, web-native) or native AI/EPS files (standard for print production). When vectorizing logos or illustrations for brand systems, always request native format output, not SVG, to preserve full editability.

Video Formats

AI video tools output MP4 (H.264/AVC or H.265/HEVC) as standard. Some offer ProRes or DNxHD for professional workflows. For social content, H.264 MP4 is sufficient. For broadcast or color-grading workflows, ProRes preserves quality through post-production.

3D Formats

AI 3D generators output OBJ (universal mesh), FBX (with animation support), GLTF/GLB (web-optimized, standard for web AR/VR), or USDZ (Apple ecosystem). For architectural workflows, OBJ or FBX import into Revit, SketchUp, or Blender.

Text & Document Formats

LLMs output plain text, Markdown, HTML, JSON, or structured document formats. For client deliverables, always convert AI-generated text into your organization's standard templates — never deliver raw AI output. JSON is essential for API integrations and data-driven creative systems.

Audio Formats

AI music generators output MP3 (compressed), WAV (uncompressed, studio-quality), or MIDI (musical note data for arrangement editing). For brand sonic identities, always request WAV masters. MIDI files allow editing in digital audio workstations.

10. Intellectual Property, Ethics & Responsible AI

Training Data Provenance

Every generative AI model was trained on a dataset of existing creative work. The legal status of this training is contested and varies by jurisdiction. In the United States, the fair use doctrine is the primary defense raised by AI companies, but this has not been definitively tested at the Supreme Court level. In the European Union, the AI Act imposes transparency requirements on training data. Be aware that AI-generated imagery may contain stylistic DNA from thousands of uncredited human artists.

Client Disclosure

Establish clear policies on when and how to disclose AI usage to clients. A recommended approach: disclose AI usage as a production technique, not a creative shortcut. For brand identity work where originality is paramount, AI should be used for exploration and inspiration, not for final deliverables. For campaign production where efficiency is valued, AI augmentation should be transparently documented.

Output Ownership

In the United States, the Copyright Office has consistently held that purely AI-generated works are not eligible for copyright protection. Human creative input — selection, arrangement, modification, curation — is required for copyrightability. This means that AI-generated concepts shared without human modification may not be protectable by intellectual property law.

Bias and Representation

AI models reproduce and often amplify biases present in their training data. Image generation models have historically underrepresented certain demographics and overrepresented Western aesthetic norms. Text models may default to masculine pronouns for leadership roles. Every AI-generated asset must be reviewed through a diversity, equity, and inclusion lens before delivery.

Deepfake and Synthetic Media

AI avatar generators, voice cloning tools, and video generation systems create synthetic media that can be indistinguishable from authentic recordings. Establish a strict policy: synthetic representations of real people require explicit written consent. This includes AI-generated images of people for campaign use — while legal, the ethical implications must be considered case by case.

11. The Creative Workflow Integration Model

Understanding individual AI concepts and tools is necessary but not sufficient. The following four-phase model provides a framework for integrating AI into creative workflows. It is discipline-agnostic and serves as a mental model for approaching any project.

Phase 1: Divergence (Idea Generation)

The goal is quantity over quality. AI tools generate the widest possible set of concepts, references, and starting points. You act as a curator, not a creator — selecting, combining, and redirecting AI output.

• Appropriate tools: Midjourney for visual concepts, GPT-5.4/Claude 4.6 for strategic frameworks, Suno/Udio for sonic mood exploration, Napkin AI for diagram generation
• Key discipline: Do not judge at this stage. Generate abundantly.

Phase 2: Convergence (Selection & Refinement)

Apply human judgment to narrow the generated set to the most promising directions. AI tools shift from generation to evaluation — predicting which concepts will resonate, analyzing competitive differentiation, or simulating consumer response.

• Appropriate tools: Perplexity for competitive validation, Atypica.AI for consumer response prediction, Maze for prototype testing
• Key discipline: Human taste and strategic judgment are irreplaceable. AI provides data; humans make choices.

Phase 3: Production (Execution at Scale)

The selected concept is executed and produced. AI tools accelerate production workflows — automating repetitive tasks, generating variations for localization, converting formats, and enabling rapid iteration.

• Appropriate tools: Adobe Firefly for generative fill, Figma AI for component scaling, Synthesia for multilingual video production, Abyssale for asset batch generation
• Key discipline: AI accelerates production but does not eliminate craft. Every AI-generated asset requires human refinement before delivery.

Phase 4: Governance (Verification & Handoff)

The work is reviewed for quality, accuracy, brand alignment, and legal compliance. AI tools assist with verification — checking accessibility, flagging potential IP concerns, and ensuring consistency.

• Appropriate tools: Grammarly AI for tone compliance, Writer.ai for terminology governance, Adobe Acrobat AI for document verification
• Key discipline: AI-generated work must pass the same quality gates as human-created work. There are no shortcuts in governance.

12. Glossary

Agentic AI / AI Agent: An autonomous system that perceives its environment, makes decisions dynamically, and selects tools or actions to achieve goals without following a predefined sequence.

Diffusion Model: A generative model that learns to reverse a gradual noising process. Starting from random noise, it iteratively applies learned denoising steps to generate coherent images, video, or audio. Examples: Stable Diffusion, DALL-E, Midjourney.

Discriminative AI: Machine learning models that learn the boundaries between categories to classify or predict labels. Does not generate new content.

Foundation Model: A large-scale neural network pretrained on broad data that serves as a base for many downstream tasks. May be unimodal or multimodal.

GAN (Generative Adversarial Network): A generative architecture with two competing networks — a generator that creates synthetic data and a discriminator that judges its authenticity.

Generative AI: Models that learn the probability distribution of training data to generate new, statistically similar content.

Large Language Model (LLM): A foundation model specialized for natural language. Examples: GPT-5.4, LLaMA 4, Mistral Large 3.

LoRA (Low-Rank Adaptation): A parameter-efficient fine-tuning technique for adapting pretrained models to specific styles or domains.

Multimodal Model: A foundation model capable of processing text, image, audio, and video within a single architecture. Examples: GPT-5.4, Gemini 3.1 Pro, Claude Opus 4.6.

Prompt Engineering: The practice of crafting input text to guide generative AI models toward desired outputs.

RAG (Retrieval-Augmented Generation): An architecture that grounds LLM outputs in external knowledge by retrieving relevant documents, reducing hallucination.

Transformer: The dominant neural network architecture for modern AI, using self-attention mechanisms to process input data.

Vibe Coding: A software development workflow where the developer describes intent in natural language and guides an AI assistant to generate, refine, and debug code.

Workflow (AI Workflow): A predefined, deterministic sequence of operations orchestrated by explicit logic. Prioritizes reliability and auditability over autonomy.

This guide is a living document. The AI landscape evolves rapidly — verify specific tool capabilities and pricing on vendor websites before making decisions.