AI for UX
Glossary
52 AI terms a designer is likely to meet, in plain language, grouped by theme.
How AI models work
| Term | In one line |
|---|---|
| Context window | How much text (in tokens) a model can consider at once: your prompt, attached files, the conversation so far and its own answer. |
| Diffusion model | The kind of model behind most image and video generators. It starts from noise and refines it step by step towards what the prompt describes. |
| Embeddings | A way to turn text (or images) into numbers so that similar meanings end up close together. It powers semantic search, clustering and recommendations. |
| Fine-tuning | Further training a model on your own examples so it adopts a style or task. Expensive and rarely needed for design work: good context usually gets you further. |
| Knowledge cutoff | The date after which a model has no training data. Without search or connected tools, it doesn't know about newer products, prices or laws. |
| Large language model (LLM) | A model trained on huge amounts of text to predict the next token. It powers chat assistants, copilots and agents; it generates plausible text, which isn't the same as true text. |
| Multimodal | A model that understands (and sometimes produces) more than text: images, screenshots, audio, video or PDFs. |
| Open-weight model | A model whose trained weights you can download and run yourself, under its own licence. Not always "open source": the training data and code may stay closed. |
| Quantisation | Compressing a model's numbers (for example to 8 or 4 bits) so it needs less memory and runs faster, with a small loss of quality. |
| Reasoning model | A model that 'thinks' through intermediate steps before answering. Better at planning, analysis and multi-step tasks; slower and more expensive. |
| Temperature | A setting that controls how random a model's output is. Low = consistent and predictable; high = more varied and creative. It's why the same prompt can give different answers. |
| Token | The unit a model reads and writes: roughly ¾ of an English word; other languages often need more tokens for the same text. Limits, speed and price are all counted in tokens. |
Prompting and context
| Term | In one line |
|---|---|
| Few-shot prompting | Giving the model a few examples of the output you want inside the prompt. Zero-shot = no examples. |
| Grounding | Tying the model's answer to specific sources (documents, data, a Figma file) instead of its general training, ideally with citations. |
| Prompt | The instruction and input you give a model. A good prompt is a good brief: role, context, task, format and constraints. |
| Retrieval-augmented generation (RAG) | The model first searches a document set (wiki, research repository, help centre) and then answers using what it found. |
| System prompt | Hidden instructions a product gives the model before the user's message: its role, tone, rules and limits. |
Agents and tools
| Term | In one line |
|---|---|
| Agent skills | Packaged instructions, scripts and references an agent loads for a specific task, e.g. 'review a screen against our library'. Reusable team know-how in a file. |
| AI agent | An AI system that pursues a goal over several steps: it plans, uses tools, checks results and continues, with more autonomy than a chat answer. |
| Copilot | AI that assists inside a tool while the human stays in control: suggestions, drafts, autocomplete. The step before an agent on the autonomy spectrum. |
| Sub-agents | Helper agents that a main agent sends off to do heavy work in their own context and report back a summary, often in parallel. |
| Tool use | Also called function calling: the model calls external tools (search, a database, Jira, Figma, a browser) instead of only producing text. |
Building with AI
| Term | In one line |
|---|---|
| Context files | Files that agents read automatically (CLAUDE.md, AGENTS.md, DESIGN.md) holding principles, DS rules and conventions. |
| Design tokens | Named design decisions (colour, spacing, type, radius) stored as data, so tools, code and AI agents use the same values. |
| Inference cost | The money, time and energy spent every time a model produces an answer. It grows with model size, tokens and how often a feature runs. |
| MCP | Model Context Protocol: an open standard for connecting AI agents to tools and data, e.g. Figma, browsers, Jira. |
| Spec-driven development | Work starts from written specs (requirements → design → tasks) that humans review and agents implement. |
| Vibe coding vs production code | Fast AI-generated code is great for prototypes and terrible as unreviewed production code. Make the boundary explicit. |
Quality and risk
| Term | In one line |
|---|---|
| Algorithmic bias | Systematically worse or unfair output for some groups, languages or cases, inherited from training data or design choices. |
| Automation bias | The tendency to over-trust automated suggestions and stop checking them, especially when tired, rushed or not an expert. |
| Error analysis | Reading real AI outputs, writing down what went wrong, grouping the problems and counting them, before writing any criteria. |
| Evals | Structured tests of AI output against defined criteria and test cases. The AI equivalent of usability testing plus QA. |
| Guardrails | Rules and checks around a model that block or redirect unsafe, off-topic or unauthorised behaviour. |
| Hallucination | When a model states something false with confidence: an invented quote, feature, statistic or source. |
| LLM-as-a-judge | Using one AI model to grade another model's answers against a rubric. Fast and scalable, but biased, so it must be checked against human grading. |
| Model drift | When an AI feature's behaviour changes over time, because the model was updated, the data changed, or users use it differently. |
| Personal data in prompts | Names, contact details, customer data or recordings pasted into an AI tool. Whether that's allowed depends on the tool's approval, data retention and training terms. |
| Prompt injection | Hidden instructions inside content the AI reads (an email, a web page, a document) that try to make it do something else, such as leak data or take an action. |
Designing AI features
| Term | In one line |
|---|---|
| AI error states | What the UI does when AI is wrong, slow, unsure, refuses, or isn't available: the empty, loading and error states of AI features. |
| AI fatigue | People tiring of AI features that add little: chat boxes where search worked, assistants nobody asked for, the same "AI" label everywhere. |
| Confidence indicators | Signals that show how sure the AI is (or how well-supported its answer is), so users know when to double-check. |
| Explainability | Helping users understand why the AI produced an answer or suggestion: sources, main factors, confidence, what it couldn't do. |
| Feedback loops | Ways users tell the system whether output was useful (thumbs, edits, corrections) so the team can improve it. |
| Generative UI | Interfaces that are partly generated at runtime by AI, e.g. a chart, form or summary built for the user's request, rather than fully designed in advance. |
| Human-in-the-loop | AI proposes, a named human decides and stays accountable, with clear review, approval and undo points. |
| Streaming responses | Showing AI output as it's generated, word by word, instead of waiting for the full answer. |
| Trust calibration | Users should trust AI exactly as much as it deserves: not blindly (automation bias), not never (wasted value). |
Research
| Term | In one line |
|---|---|
| Convenience sample | Survey or study participants who were easy to reach or chose to answer, rather than a representative sample. Useful, but it can't speak for everyone. |
| Synthetic users | AI-simulated 'users' answering research questions. Useful for brainstorming hypotheses, never a substitute for real users. |
Adopting AI in the team
| Term | In one line |
|---|---|
| AI adoption ladder | Explore → Assist → Co-create → Orchestrate: where a designer is with AI for a given task. |
| AI champions network | One champion per 3–5 designers tries tools, writes playbook entries and runs office hours. Rotate the role. |
| AI usage guidelines | A one-page 'how we use AI in UX': approved tools, data rules, review rules, transparency and where to ask. |
All terms, A to Z
Agent skills
Packaged instructions, scripts and references an agent loads for a specific task, e.g. 'review a screen against our library'. Reusable team know-how in a file.
Why it matters for UX
- Turn a repeated checklist or review routine into a skill
- Keep skills small and specific; version them like a design asset
- Share them across the team as a plugin
AI adoption ladder
Explore → Assist → Co-create → Orchestrate: where a designer is with AI for a given task.
How to use it
- Place yourself (or a teammate) per task, not overall
- Move one rung at a time
- On low-rung tasks, ask for pairing or a worked example
Learn more
AI agent
An AI system that pursues a goal over several steps: it plans, uses tools, checks results and continues, with more autonomy than a chat answer.
Why it matters for UX
- Decide how much autonomy it gets and where a human approves
- Design for visibility: plan preview, progress, what it changed, undo
- Start read-only; add write actions once trust is earned
AI champions network
One champion per 3–5 designers tries tools, writes playbook entries and runs office hours. Rotate the role.
How to use it
- Pick enthusiasts and one thoughtful sceptic
- Give them time (e.g. 10%)
- Champions curate the prompt library
Learn more
AI error states
What the UI does when AI is wrong, slow, unsure, refuses, or isn't available: the empty, loading and error states of AI features.
Why it matters for UX
- Design them in the spec, like any other state
- Always give a manual path to finish the task
- Use plain language: what happened, what the user can do
AI fatigue
People tiring of AI features that add little: chat boxes where search worked, assistants nobody asked for, the same "AI" label everywhere.
Why it matters for UX
- NN/g called 2026 "the year of AI fatigue"
- Start from the user's problem, not from "where can we add AI?"
- A good non-AI path is often the better feature
AI usage guidelines
A one-page 'how we use AI in UX': approved tools, data rules, review rules, transparency and where to ask.
How to use it
- Check it before trying a new tool or sharing data
- Co-write with IT security & DPO
- One page; link the details
- Review every quarter
Learn more
Algorithmic bias
Systematically worse or unfair output for some groups, languages or cases, inherited from training data or design choices.
Why it matters for UX
- Test with diverse inputs: languages, regions, roles, edge cases
- Watch generated personas and images for stereotypes
- Treat fairness as part of quality, not an extra
Automation bias
The tendency to over-trust automated suggestions and stop checking them, especially when tired, rushed or not an expert.
Why it matters for UX
- Design reviews and approvals so people actually check, not just click OK
- Show uncertainty and sources
- In research, watch whether users accept wrong suggestions
Confidence indicators
Signals that show how sure the AI is (or how well-supported its answer is), so users know when to double-check.
Why it matters for UX
- Prefer meaningful signals (source count, 'needs review') over fake-precise percentages
- Make low confidence change the flow: ask, flag or hand over
- Test whether users actually notice and understand them
Context files
Files that agents read automatically (CLAUDE.md, AGENTS.md, DESIGN.md) holding principles, DS rules and conventions.
How to use it
- Keep them short and specific
- Write 'never' rules explicitly
- Version and review them like code
Learn more
- CLAUDE.md / memory (Claude Code docs)
- AGENTS.md
- DESIGN.md format (Google Labs)
- Context Libraries (NN/g)
- UX-Context Design (NN/g)
Context window
How much text (in tokens) a model can consider at once: your prompt, attached files, the conversation so far and its own answer.
Why it matters for UX
- Long chats drift: start a fresh conversation for a new task
- Put the important instructions where they won't get lost: in context files or at the start
- Big inputs (a 60-minute transcript) are a reason to use sub-agents or summaries
Convenience sample
Survey or study participants who were easy to reach or chose to answer, rather than a representative sample. Useful, but it can't speak for everyone.
Why it matters for UX
- Most AI-in-design surveys are convenience samples: read their figures as signals
- Ask who answered and who didn't before quoting a percentage
- Vendor surveys sample their own users
Copilot
AI that assists inside a tool while the human stays in control: suggestions, drafts, autocomplete. The step before an agent on the autonomy spectrum.
Why it matters for UX
- Good default for expert users: speeds them up without taking over
- Make suggestions easy to accept, edit or ignore
- Measure whether suggestions are accepted, edited or rejected
Design tokens
Named design decisions (colour, spacing, type, radius) stored as data, so tools, code and AI agents use the same values.
How to use it
- Single source of truth in the DS
- Reference tokens in specs and CLAUDE.md
- Lint for raw values
Learn more
Diffusion model
The kind of model behind most image and video generators. It starts from noise and refines it step by step towards what the prompt describes.
Why it matters for UX
- Great for moodboards, illustrations and concept visuals; poor at exact text, logos and UI details
- Check licensing and brand rules before using generated images
- Results vary run to run: generate several and curate
Embeddings
A way to turn text (or images) into numbers so that similar meanings end up close together. It powers semantic search, clustering and recommendations.
Why it matters for UX
- It's what lets search find 'delivery delay' when the text says 'shipment late'
- Useful for clustering research notes or support tickets by theme
- Clusters still need a human to name and judge them
Error analysis
Reading real AI outputs, writing down what went wrong, grouping the problems and counting them, before writing any criteria.
Why it matters for UX
- It is qualitative analysis, the same skill as synthesising usability findings
- The biggest failure category tells you what to fix first
- Criteria written after reading are sharper than criteria guessed up front
Evals
Structured tests of AI output against defined criteria and test cases. The AI equivalent of usability testing plus QA.
Why it matters for UX
- UX can write the test cases: real tasks, edge cases, tricky inputs
- Define 'good' in plain language before measuring it
- Re-run evals when the prompt, model or data changes
Explainability
Helping users understand why the AI produced an answer or suggestion: sources, main factors, confidence, what it couldn't do.
Why it matters for UX
- Explain at the level the user needs, not the model internals
- Short 'why' plus a link to sources beats a long disclaimer
- Required for some decisions under the EU AI Act
Feedback loops
Ways users tell the system whether output was useful (thumbs, edits, corrections) so the team can improve it.
Why it matters for UX
- Edits and corrections are richer signals than thumbs up/down
- Ask for feedback at the moment of use, not in a survey later
- Close the loop: show users their feedback changed something
Few-shot prompting
Giving the model a few examples of the output you want inside the prompt. Zero-shot = no examples.
Why it matters for UX
- Show 2–3 good examples of a microcopy pattern, a finding or a ticket comment
- Examples steer more than adjectives ('concise, friendly')
- Vary your examples, or it will copy them too closely
Fine-tuning
Further training a model on your own examples so it adopts a style or task. Expensive and rarely needed for design work: good context usually gets you further.
Why it matters for UX
- Try better prompts, context files or RAG first
- If someone proposes fine-tuning, ask what data it uses and who approved it
- A fine-tuned model still needs evals
Generative UI
Interfaces that are partly generated at runtime by AI, e.g. a chart, form or summary built for the user's request, rather than fully designed in advance.
Why it matters for UX
- Design the system (components, rules, tokens) rather than every screen
- Define what the AI may and may not generate
- Keep layouts predictable for expert users
Grounding
Tying the model's answer to specific sources (documents, data, a Figma file) instead of its general training, ideally with citations.
Why it matters for UX
- Ask for sources or timestamps with every claim
- Give it the actual spec, transcript or page instead of asking from memory
- Ungrounded answers are where hallucinations hide
Guardrails
Rules and checks around a model that block or redirect unsafe, off-topic or unauthorised behaviour.
Why it matters for UX
- Design the guardrail message: say what can't be done, why, and what to do instead
- Guardrails aren't perfect: keep a human check for high-risk actions
- Test them with adversarial and edge-case inputs
Hallucination
When a model states something false with confidence: an invented quote, feature, statistic or source.
Why it matters for UX
- Spot-check quotes, numbers and links against the source
- Ask for citations; a missing or fake citation is a warning sign
- In AI features, design for it: sources, 'verify' prompts, easy correction
Human-in-the-loop
AI proposes, a named human decides and stays accountable, with clear review, approval and undo points.
How to use it
- Name the decision owner
- Design explicit approval steps for risky actions
- Make it easy to correct the AI
Learn more
- Guidelines for Human-AI Interaction (Amershi et al., CHI 2019)
- Designing for Agentic AI: Control, Consent, Accountability (Smashing, 2026)
Inference cost
The money, time and energy spent every time a model produces an answer. It grows with model size, tokens and how often a feature runs.
Why it matters for UX
- A rule or a search often does the same job for a fraction of the cost
- Long prompts and reasoning steps cost tokens on every call
- Cost shapes design choices: when to call the model, how much context to send
Knowledge cutoff
The date after which a model has no training data. Without search or connected tools, it doesn't know about newer products, prices or laws.
Why it matters for UX
- Ask it to search or give it the current source when facts matter
- Be sceptical of answers about recent tool features (including Figma's)
- Date-stamp facts in your own notes, as this guide does
Large language model (LLM)
A model trained on huge amounts of text to predict the next token. It powers chat assistants, copilots and agents; it generates plausible text, which isn't the same as true text.
Why it matters for UX
- Treat output as a fluent draft, not a source of truth
- Expect it to be good at language, structure and patterns; weaker at exact facts, maths and up-to-date data
- Different models have different strengths, costs and data terms: the choice matters
LLM-as-a-judge
Using one AI model to grade another model's answers against a rubric. Fast and scalable, but biased, so it must be checked against human grading.
Why it matters for UX
- Known biases: favouring the first answer shown, longer answers and its own style
- Grade 20 to 30 answers yourself and compare before relying on it
- The rubric it applies is design work
MCP
Model Context Protocol: an open standard for connecting AI agents to tools and data, e.g. Figma, browsers, Jira.
How to use it
- Figma MCP: read frames, variables, Code Connect
- Playwright MCP: screenshots & checks
- Only connect approved servers
Learn more
Model drift
When an AI feature's behaviour changes over time, because the model was updated, the data changed, or users use it differently.
Why it matters for UX
- A feature that tested well can degrade silently
- Plan regular re-checks (evals) and watch feedback signals
- Note the model version in research and test reports
Multimodal
A model that understands (and sometimes produces) more than text: images, screenshots, audio, video or PDFs.
Why it matters for UX
- Share screenshots of a design for critique or to describe a bug
- Check what a tool does with uploaded images before using real customer data
- Image understanding is good but not pixel-perfect: verify measurements
Open-weight model
A model whose trained weights you can download and run yourself, under its own licence. Not always "open source": the training data and code may stay closed.
Why it matters for UX
- Runs on your own device, so prompts and files can stay local
- Each model has its own licence: read it before commercial use
- Usually smaller and weaker than the largest cloud models
Personal data in prompts
Names, contact details, customer data or recordings pasted into an AI tool. Whether that's allowed depends on the tool's approval, data retention and training terms.
Why it matters for UX
- Default: anonymise or use fixtures before prompting
- Check whether the tool keeps your data or trains on it
- Interview recordings and live-system screenshots count as personal data
Prompt
The instruction and input you give a model. A good prompt is a good brief: role, context, task, format and constraints.
Why it matters for UX
- Keep stable context in context files; keep the prompt for the task
- Say what 'good' looks like and what to avoid
- Iterate: refine the prompt instead of fixing every output by hand
Prompt injection
Hidden instructions inside content the AI reads (an email, a web page, a document) that try to make it do something else, such as leak data or take an action.
Why it matters for UX
- Never let an agent act on untrusted content without a human check
- Limit what tools and data an agent can reach
- Treat any 'the AI did something odd' report seriously
Quantisation
Compressing a model's numbers (for example to 8 or 4 bits) so it needs less memory and runs faster, with a small loss of quality.
Why it matters for UX
- It is what lets a capable model run on a laptop
- 8-bit is usually indistinguishable; 4-bit trades a little quality for a lot of memory
- Model names often say it: "4bit", "q8"
Reasoning model
A model that 'thinks' through intermediate steps before answering. Better at planning, analysis and multi-step tasks; slower and more expensive.
Why it matters for UX
- Use it for synthesis, critique and planning, not for quick rewrites
- Expect longer waits: design the UI for it (progress, plan preview)
- Its visible reasoning is a summary, not proof that the answer is right
Retrieval-augmented generation (RAG)
The model first searches a document set (wiki, research repository, help centre) and then answers using what it found.
Why it matters for UX
- It's how 'ask our Confluence' assistants work
- Answers are only as good as the documents: outdated pages give outdated answers
- Design the UI to show which sources were used
Spec-driven development
Work starts from written specs (requirements → design → tasks) that humans review and agents implement.
How to use it
- UX owns the experience sections
- Specs are versioned with the code
- Review specs before generation
Learn more
Streaming responses
Showing AI output as it's generated, word by word, instead of waiting for the full answer.
Why it matters for UX
- Reduces perceived waiting; design a clear 'still working' and a stop button
- Don't let users act on half-finished output
- For agents, stream progress steps, not just text
Sub-agents
Helper agents that a main agent sends off to do heavy work in their own context and report back a summary, often in parallel.
Why it matters for UX
- Useful for long reading: transcripts, many pages, big codebases
- You only see their summary: ask for evidence (quotes, links) in it
- Parallel work is faster but costs more
Synthetic users
AI-simulated 'users' answering research questions. Useful for brainstorming hypotheses, never a substitute for real users.
How to use it
- Label outputs as hypotheses
- Never cite them as evidence
- Validate with real users
Used in: AI in UX research · AI tools landscape for designers
Learn more
System prompt
Hidden instructions a product gives the model before the user's message: its role, tone, rules and limits.
Why it matters for UX
- When designing an AI feature, the system prompt is part of the design: review it like copy
- Tone, refusals and safety behaviour mostly live here
- Assume users may try to override it
Temperature
A setting that controls how random a model's output is. Low = consistent and predictable; high = more varied and creative. It's why the same prompt can give different answers.
Why it matters for UX
- Don't expect identical output twice: review every run
- For tests and comparisons, keep inputs fixed and run more than once
- For ideation, variety is a feature: ask for several options
Token
The unit a model reads and writes: roughly ¾ of an English word; other languages often need more tokens for the same text. Limits, speed and price are all counted in tokens.
Why it matters for UX
- Long documents, transcripts and images all consume tokens
- Non-English text usually costs more tokens for the same content
- When a tool says a limit was hit, it's usually a token limit
Tool use
Also called function calling: the model calls external tools (search, a database, Jira, Figma, a browser) instead of only producing text.
Why it matters for UX
- It's what turns a chat model into something that can act
- Each tool is a permission: read-only vs. write matters
- Show users which tools an assistant used to reach its answer
Trust calibration
Users should trust AI exactly as much as it deserves: not blindly (automation bias), not never (wasted value).
How to use it
- Show confidence and sources
- Expose failures honestly
- Test with deliberately wrong suggestions
Learn more
- People + AI Guidebook (Google PAIR)
- Crafting AI Explanations for Every Role in Your Enterprise (NN/g)
- Explainable AI in Chat Interfaces (NN/g)
Vibe coding vs production code
Fast AI-generated code is great for prototypes and terrible as unreviewed production code. Make the boundary explicit.
How to use it
- Label prototypes clearly
- Separate prototype repos
- Engineering decides what moves to production
Learn more
Terms from the lessons, A to Z
The terms the lessons define in their vocabulary tables. The general glossary has those of every course.
- A2A
- The protocol that connects agents to other agents In Conversational and generative UI
- Acceptance rate
- Share of AI suggestions people use In Measuring AI features
- ADKAR
- Awareness, Desire, Knowledge, Ability, Reinforcement: five stages of individual change In Guiding designers through AI adoption
- Affinity mapping
- Grouping notes by similarity to find themes In AI-assisted synthesis, step by step
- AG-UI
- An open protocol for notices between agents and interfaces (2026 example) In Conversational and generative UI
- Agent Skills
- The open format for skills, used by several tools In Build your first skill
- AI agent
- A system that pursues a goal by planning and taking actions with tools In Agentic design
- AI exposure
- How much of a job's tasks AI could do or speed up In Your design career in the AI era
- AI fatigue
- People tiring of AI features that add little In When not to use AI
- AI literacy
- The skills and knowledge to use AI well and know its risks In Choosing and piloting AI tools
- AI-moderated interview
- An interview run by an AI system instead of a person In AI in UX research
- Algorithmic bias
- Systematic unfairness in an AI system's outputs In Ethics, bias and inclusion in AI design
- Alt text
- A text alternative that conveys an image's meaning In Accessibility with and for AI
- Anthropomorphism
- Giving a system human traits (a name, feelings, a face) In Ethics, bias and inclusion in AI design
- Giving software human traits: feelings, a name, a personal life In Write the voice of an AI assistant
- ARIA
- Attributes that tell assistive technology what an element is and does In Accessibility with and for AI
- Audit trail
- A record of what was done, by whom or what, and why In Agentic design
- Automation bias
- Trusting a system's suggestion more than you should In Ethics, bias and inclusion in AI design
- Autonomy level
- How much the AI does without asking In Agentic design
- Axial coding
- Grouping the notes into categories In Write an eval set
- Baseline
- How the task performs today, before the tool In Choosing and piloting AI tools
- Calibration
- Comparing an agent's output with past human work before trusting it In Building an AI agent roadmap for your UX team
- Candidate
- A piece of work that an agent might take over, written in one line In Building an AI agent roadmap for your UX team
- Champion
- A designer who helps others adopt a tool and collects what works In Guiding designers through AI adoption
- Change management
- Helping people move from the old way of working to the new one In Guiding designers through AI adoption
CLAUDE.md- The project's instructions for Claude Code In Generating design with Claude Code
- Code Connect
- Figma's link between a design component and its code In AI-ready design systems
- Figma's mapping from design components to code components In Generating design with Claude Code
- Codebook
- The list of codes and what each means In AI-assisted synthesis, step by step
- Coding (tagging)
- Labelling parts of the data with themes or categories In AI-assisted synthesis, step by step
- Cohort
- A group that learns together, on the same schedule In Guiding designers through AI adoption
- Confidence indicators
- Ways of showing how sure the AI is In Designing for uncertainty
- Content Credentials (C2PA)
- Attached records of how a piece of media was made In AI for visual design: images and video
- Content training
- A vendor using your files or prompts to improve its models In Choosing and piloting AI tools
- Context file
- Standing instructions an AI tool reads in every session In AI-ready design systems
- Standing instructions an agent reads in every session In UX agents and skills: a guide for designers
- Context files
- Files a tool reads every time: rules and context you don't want to repeat In Prompt patterns for designers
- Context window
- How much text the model can consider at once In Run a model on your laptop
- Controlled experiment
- Comparing groups with and without the feature, assigned fairly In Measuring AI features
- Convenience sample
- Survey answers from whoever chose to respond In Your design career in the AI era
- Copyright
- The right to control copies of an original creative work In Copyright and AI-generated assets
- Correction rate
- Share of AI suggestions people edit or undo In Measuring AI features
- Credits / tokens
- How these tools meter usage; each generation costs some In Prototyping with AI
- Dark pattern
- An interface designed to trick people into choices they wouldn't otherwise make In Ethics, bias and inclusion in AI design
- Data processing agreement (DPA)
- The contract under which a vendor processes your data In AI governance and data protection
- Data protection officer (DPO)
- The person responsible for data protection in an organisation In AI governance and data protection
- Deepfake
- Realistic AI content showing real people, places or events that didn't happen In AI for visual design: images and video
- Deployer
- Under the AI Act, an organisation that uses an AI system In AI governance and data protection
- Descriptive code
- A label for what was said ("mentions price") In AI-assisted synthesis, step by step
- Descriptive vs causal
- Showing a pattern vs proving what caused it In Your design career in the AI era
- Design critique
- A structured review by peers, against the goal and principles In AI design review process
- Design token
- A named design decision (colour, spacing, type) stored as data In AI-ready design systems
- Design-system drift
- Small deviations that slowly break consistency In AI design review process
- Deterministic
- Gives the same output for the same input, every time (rules, search, forms) In When not to use AI
- Diffusion model
- A model that creates images by removing noise step by step In AI for visual design: images and video
- A model that turns noise into an image, step by step In Generate images locally
- Digital twin
- A synthetic respondent built from your own customer or research data In Synthetic users and digital twins
- Disclosure
- Telling people that content was made by AI In AI for content design
- Telling people when and how AI was involved in your work In Your design career in the AI era
- Discussion guide
- The plan of questions and topics for an interview In AI in UX research
- DPIA
- A data protection impact assessment, done before risky processing In AI governance and data protection
- DTCG format
- The community format for tokens, stable version 2025.10 In AI-ready design systems
- Entry-level hiring
- Hiring of people early in their careers In Your design career in the AI era
- Error analysis
- Reading outputs, naming failures and counting them In Write an eval set
- Errors introduced
- Mistakes caused by AI that wouldn't otherwise happen In Measuring AI features
- Evals
- Structured tests of AI output against criteria In Evals for designers
- Explainability
- Making clear why the AI produced a result In Designing for uncertainty
- Explicit / implicit feedback
- Ratings people give, vs signals from what they do In Measuring AI features
- Fair use / exceptions
- Legal room to use protected works without permission in some cases In Copyright and AI-generated assets
- Fallback
- The non-AI path people can always take In When not to use AI
- Few-shot prompting
- Including a few examples of the output you want In Prompt patterns for designers
- Fidelity trap
- When polish makes people accept a design without questioning it In Prototyping with AI
- Frontmatter
- The block between
---lines at the top, withnameanddescriptionIn Build your first skill - Generative UI
- Interface pieces the AI composes on the fly In Conversational and generative UI
- Graceful failure
- Failing in a way that still leaves the person a good next step In Designing for uncertainty
- Hand-over
- Passing the conversation to a person or a better channel In Write the voice of an AI assistant
- Handoff / export
- How the work leaves the tool: a link, layers in Figma, code in GitHub In Prototyping with AI
- HEART
- Happiness, Engagement, Adoption, Retention, Task success In Measuring AI features
- Held-out set
- Cases kept aside, never used to tune prompts In Evals for designers
- High-risk AI
- Uses listed by the AI Act that need strict safeguards In AI governance and data protection
- Human-in-the-loop
- A person reviews or approves at key points In Agentic design
- Indemnity
- A vendor's promise to defend you, and pay, if a covered claim arises In Copyright and AI-generated assets
- Inference cost
- The money and energy spent every time a model answers In When not to use AI
- Interpretive code
- A label for what it means ("feels judged by the app") In AI-assisted synthesis, step by step
- Interrupt
- A pause where the agent waits for a person's decision In Conversational and generative UI
- Inventory
- A rough map of what the team actually spends time on In Building an AI agent roadmap for your UX team
- Licence
- The terms that say what you may do with the model and its output In Generate images locally
- Likeness
- A person's recognisable face, voice or identity In Copyright and AI-generated assets
- Live region
- A part of the page whose changes are announced automatically In Accessibility with and for AI
- LLM-as-a-judge
- Using an AI model to grade AI answers In Evals for designers
- Localisation
- Adapting content and product to a culture, not only translating it In AI for content design
- MCP
- A standard way for AI tools to connect to other apps and data In AI-ready design systems
- The protocol that connects agents to tools and data In Conversational and generative UI
- The standard that connects the agent to tools like Figma or a browser In Generating design with Claude Code
- The protocol that connects agents to other systems In UX agents and skills: a guide for designers
- Mode collapse
- When a model's answers bunch around a few typical options In Synthetic users and digital twins
- N-best alternatives
- Showing the top few options instead of one answer In Designing for uncertainty
- Notice (event)
- A small message an agent sends to say what it's doing In Conversational and generative UI
- Now-Next-Later roadmap
- A roadmap by phase instead of by date In Building an AI agent roadmap for your UX team
- Open coding
- Writing free notes on each output, without categories yet In Write an eval set
- Open model
- A model whose weights you can download and run yourself In Run a model on your laptop
- Open weights
- A model you can download and run yourself, under its licence In AI for visual design: images and video
- Over-trust / under-trust
- Relying on AI more, or less, than it deserves In Designing for uncertainty
- Parameters
- The numbers a model learned; "4B" means about 4 billion In Run a model on your laptop
- Pass rate
- The share of cases that meet a criterion In Write an eval set
- Personal data
- Any information about an identifiable person In AI governance and data protection
- Pilot
- A short, planned trial on real work, with a decision at the end In Choosing and piloting AI tools
- Plan mode
- The agent proposes before it changes anything In Generating design with Claude Code
- Plan preview
- Showing the steps an agent intends to take, before it acts In Agentic design
- Playbook
- A shared collection of proven ways to do a task In Guiding designers through AI adoption
- Plugin
- A bundle of skills, agents and settings, installed in one step In UX agents and skills: a guide for designers
- Probabilistic
- Can give different outputs for the same input (AI models) In When not to use AI
- Prompt
- Everything you give the model for a task In Prompt patterns for designers
- Prompt injection
- Hidden instructions that try to make the AI ignore its own In Write the voice of an AI assistant
- Prompt-to-prototype
- Tools that build a working app from a description In Prototyping with AI
- Protocol
- An agreed format so different systems can talk to each other In Conversational and generative UI
- Pseudonymisation
- Replacing names with codes; still personal data In AI-assisted synthesis, step by step
- Quantised (4-bit)
- Compressed so it needs less memory, with a small loss of quality In Run a model on your laptop
- Quantised (8-bit)
- Compressed to use less memory, with little visible loss In Generate images locally
- Reasoning ("thinking")
- Text the model writes for itself before it answers In Run a model on your laptop
- Reference image
- Your image given to the model as a guide for style or subject In AI for visual design: images and video
- References
- Extra files the skill reads only when needed In Build your first skill
- Registry
- A shared catalogue of components that tools can install from In AI-ready design systems
- Regression
- Something that used to work and now doesn't In Evals for designers
- Repository (repo)
- The folder of code and its history, usually on GitHub In Generating design with Claude Code
- Review gate
- A checkpoint work must pass before moving on In AI design review process
- Rubric
- Written criteria for grading an answer In Evals for designers
- Sandbox
- A safe test space where mistakes don't matter In UX agents and skills: a guide for designers
- Screen reader
- Software that reads the screen aloud or in braille In Accessibility with and for AI
- Screener
- Questions that check whether someone fits the study In AI in UX research
- Seed
- The number that fixes the starting noise, so a result can be reproduced In AI for visual design: images and video
- The number that fixes the randomness, so you can repeat an image In Generate images locally
- Self-review
- The maker checks their own work first, against a checklist In AI design review process
- Silicon sample
- A set of AI respondents used in place of a survey sample In Synthetic users and digital twins
- Skill
- A folder of instructions an AI tool loads for one kind of task In Build your first skill
- A reusable package of instructions for one kind of task. See Agent skills In Generating design with Claude Code
- Packaged instructions for one kind of task In UX agents and skills: a guide for designers
SKILL.md- The skill's main file: frontmatter plus instructions In Build your first skill
- Steps
- How many refinement passes the model makes In Generate images locally
- Sub-agent
- A separate AI worker with its own context, returning a summary In UX agents and skills: a guide for designers
- Sycophancy
- The tendency of a model to agree with and flatter the person asking In Synthetic users and digital twins
- Flattering or agreeing with the user instead of being accurate In Write the voice of an AI assistant
- Synthesis
- Turning raw research data into findings and insights In AI in UX research
- Synthetic users
- AI-generated respondents that imitate users In AI in UX research
- AI-generated respondents that imitate a kind of person In Synthetic users and digital twins
- System prompt
- Hidden instructions the tool adds before your prompt In Prompt patterns for designers
- Standing instructions a product gives the model before every conversation In Write the voice of an AI assistant
- Temperature
- How much the answers vary between attempts In Prompt patterns for designers
- Terminology / glossary
- The approved words for product concepts, per language In AI for content design
- Text and data mining
- Automated analysis of large amounts of content, including for AI training In Copyright and AI-generated assets
- Tokens per second
- How fast the model writes In Run a model on your laptop
- Tone
- How the voice adapts to a moment (an error vs a success) In AI for content design
- Total time
- Time to a usable result, including prompting, checking and editing In Choosing and piloting AI tools
- Trade mark
- A protected sign (a name, logo) that identifies a brand In Copyright and AI-generated assets
- Transcript
- The written record of what was said In AI in UX research
- Trust calibration
- Trust that matches how reliable the system really is In Designing for uncertainty
- Helping people trust a system exactly as much as it deserves In Ethics, bias and inclusion in AI design
- UX debt
- Problems shipped now that cost more to fix later In AI design review process
- Vendor lock-in
- When leaving a tool becomes too costly or hard In Choosing and piloting AI tools
- Vibe coding
- Building software by describing it to AI without reviewing the code. See Vibe coding vs production code In Prototyping with AI
- Voice
- Your product's personality, stable across all copy In AI for content design
- Voice and tone
- Voice is who the assistant is; tone is how it adjusts to the moment In Write the voice of an AI assistant
- WCAG
- The Web Content Accessibility Guidelines, the international standard In Accessibility with and for AI
- Weighted scoring
- Scoring options on several criteria, some counting more In Building an AI agent roadmap for your UX team