Glossary

AI terms for learning designers

56 terms in plain words. Each term has three depths: what you need to use it, how to explain it, and what AI teams talk about. The ID bridge links it to an idea you already know.

.docx file

Also called: Word file

The file format Microsoft Word uses, which is a zip folder of text files and images.

Use it
Comments, headers, and alt text are stored in separate parts, so use a tool that reads all of them.
Explain it
Tools can open the zip and read each part.
Build a career on it
Document parsing is a common first step in AI content pipelines.
ID bridge
Like a course package that holds many files inside one file.

agent

Also called: AI agent

An AI that can use tools, such as reading files or running commands, and repeat steps until a task is done.

Use it
An agent can change files on your computer, so you choose the folder and approve its actions.
Explain it
Agents work in a loop: plan, act with a tool, look at the result, and continue. Permission settings limit what they can do alone.
Build a career on it
Agent design covers tool use, orchestration of subagents, and evaluating whole task paths, not just single answers.
ID bridge
Like a production assistant who can open files and use software, not only give advice.

AI model

Also called: model

The trained program that does the AI work, for example Claude Opus or Claude Sonnet.

Use it
Different models trade speed, cost, and quality. The default model is fine for this course.
Explain it
Larger models usually reason better but are slower and use more of your plan's limits.
Build a career on it
Teams compare models with evaluations on their own tasks, not only public benchmarks.
ID bridge
Like choosing between a senior and a junior writer for a task.

AI upscaling

Also called: AI upscaler, super-resolution

Making an image larger with an AI model that predicts new detail.

Use it
It looks sharper but can change small text, numbers, and lines. Compare with the original.
Explain it
The model was trained on pairs of small and large images and guesses what the missing detail should look like.
Build a career on it
Super-resolution models (for example GAN-trained ones like Real-ESRGAN) are judged on both sharpness and faithfulness.
ID bridge
Like asking an artist to redraw a blurry diagram from memory.

AI-use disclosure

Also called: disclosure

A short statement in a document that says how AI was used and that people reviewed the result.

Use it
Say what AI did (drafting, image enhancement) and who reviewed and approved the content.
Explain it
Many organizations now require disclosure in their AI policies.
Build a career on it
Disclosure and content provenance (for example content credentials) are growing governance topics.
ID bridge
Like crediting sources and reviewers in a course's front matter.

alt text

Also called: alternative text

A short text description of an image for people who cannot see it.

Use it
Describe what the learner needs from the image, including key values. Do not start with "image of".
Explain it
Decorative images get empty alt text so screen readers skip them.
Build a career on it
Writing and evaluating alt text is a common accessibility and AI-evaluation task.
ID bridge
Like the narration for a visual in an e-learning module.

anchoring

Also called: anchoring bias

When the first number or opinion you see pulls your own judgment toward it.

Use it
Score the draft yourself before you read Claude's review.
Explain it
Independent first scores make calibration meaningful.
Build a career on it
Evaluation workflows hide other raters' scores to avoid anchoring.
ID bridge
Like grading before reading the previous teacher's comments.

automation bias

The habit of trusting a machine's output too much, even when it is wrong.

Use it
Keep checking AI outputs, especially when they look polished.
Explain it
Polished, confident output lowers people's guard. Checklists and evidence requirements help.
Build a career on it
Human-AI teaming research studies how to reduce automation bias.
ID bridge
Like accepting a spellchecker suggestion that changes the meaning.

benchmark

Also called: benchmarks

A standard public test used to compare AI models.

Use it
A high benchmark score does not guarantee good results on your own tasks. Test with your content.
Explain it
Benchmarks measure narrow skills; your work may need different ones.
Build a career on it
Teams combine public benchmarks with their own evaluations.
ID bridge
Like a standardized test versus a performance assessment on the job.

calibration

Also called: rater calibration

Raters score the same examples, compare, and discuss until they apply the rubric the same way.

Use it
Compare your scores with Claude's review and the reference review, and look at the evidence behind any difference.
Explain it
Calibration improves agreement and makes scores more trustworthy.
Build a career on it
Calibration sessions and agreement targets are standard in evaluation teams.
ID bridge
Like grader moderation meetings.

chatbot

Also called: chat assistant

An AI you talk with in a chat window that answers but usually cannot work on your files directly.

Use it
Good for questions and drafts. For work on many files, an agent is more useful.
Explain it
Chat interfaces can add tools, such as file uploads and code execution, so the line between chatbot and agent is blurring.
Build a career on it
Conversation design and evaluation of chat quality are separate job areas.
ID bridge
Like asking a colleague for advice versus asking them to do the task.

classic upscaling

Also called: Lanczos, resampling

Making an image larger by spreading the pixels it already has, without adding new detail.

Use it
It looks softer but does not invent anything, so it is safer for technical images.
Explain it
Methods like Lanczos resampling calculate new pixels from neighboring ones.
Build a career on it
Classic resampling is the baseline when evaluating AI upscalers.
ID bridge
Like enlarging a photocopy: blurrier, but nothing new is added.

Claude Code

Anthropic's agent that works in a folder on your computer, available in the Code tab of the Claude desktop app and in a terminal.

Use it
You give it a task in plain language; it reads files, runs tools, and shows you changes to approve.
Explain it
It includes project memory (CLAUDE.md), skills, subagents, and connectors. It needs a paid Claude plan.
Build a career on it
Coding agents are now used for many non-coding tasks, which is why learning teams use them.
ID bridge
A production workbench that follows your standards.

commercial account

Also called: business plan, Team plan, Enterprise plan

An organization's Claude account (Team or Enterprise, or the API) covered by business terms.

Use it
If your organization has one, it is usually the approved place for work documents.
Explain it
Under commercial terms, Anthropic does not train models on your data by default.
Build a career on it
Enterprise AI rollouts include data policies, admin controls, and training for staff.
ID bridge
Like the company-licensed version of a tool, with IT controls.

connector (MCP)

Also called: MCP, connectors

A link that lets Claude use another app, such as a calendar or a document system, through a standard called MCP.

Use it
Only connect tools your organization approves, and check what access you give.
Explain it
MCP (Model Context Protocol) is an open standard for connecting AI tools to data and apps.
Build a career on it
Building and securing connectors is part of AI integration work.
ID bridge
Like giving a new team member access to one shared drive, not all of them.

consumer account

Also called: personal plan

A personal Claude account (Free, Pro, or Max) covered by consumer terms.

Use it
Do not put company documents into a personal account unless your organization allows it.
Explain it
Consumer accounts have a setting that controls whether chats can be used to improve models.
Build a career on it
Knowing the difference between consumer and commercial terms is basic AI governance.
ID bridge
Like using your personal email for work files: usually not allowed.

context window

Also called: context

The text an AI model can see and use at one time, including your files and the conversation.

Use it
The AI only works with what is in front of it right now. Very long inputs can push earlier parts out.
Explain it
Text is split into tokens, and each model has a limit. Long sessions get summarized to make room.
Build a career on it
Context engineering, retrieval, compaction, and token budgets are daily topics on AI teams.
ID bridge
Like working memory in cognitive load theory.

data retention

How long a company keeps the data you send to its service.

Use it
Check your plan's privacy setting and your organization's rules before using real documents.
Explain it
For Claude consumer plans, retention depends on the model-improvement setting; commercial plans follow business terms.
Build a career on it
Vendor data retention and zero-data-retention options are part of AI procurement reviews.
ID bridge
Like records-retention rules for learner data in your LMS.

diff

Also called: diff view

A view that shows exactly which lines a change adds, removes, or edits in a file.

Use it
Read the diff before you accept an edit. Look hardest at numbers, units, and safety words.
Explain it
A diff compares two versions of a file line by line. Added lines and removed lines are marked, so you can review only what changed.
Build a career on it
Engineers review every change as a diff before it is merged. AI teams review model and prompt changes the same way.
ID bridge
Like tracked changes in a Word document that an SME reviews before sign-off.

evaluation (eval)

Also called: eval, evals

A planned test of how well an AI model or tool does on specific tasks.

Use it
Your rubric review of the draft is a small evaluation.
Explain it
Good evaluations use real tasks, clear scoring, and enough examples to see patterns.
Build a career on it
Writing evals is a core AI-industry skill that draws on assessment design.
ID bridge
Like a pilot test of a course with a scoring rubric.

fine-tuning

Extra training of an existing AI model on specific examples to change how it behaves.

Use it
You do not need to fine-tune a model for this course; skills and project memory are enough.
Explain it
Fine-tuning changes the model itself; skills and prompts change only what it is told.
Build a career on it
Supervised fine-tuning (SFT) data is often written by subject experts and writers.
ID bridge
Like retraining a person versus giving them a job aid.

generative AI

Also called: GenAI

AI that creates new content, such as text, images, audio, or video.

Use it
Treat its output as a draft. It can create details that are not true.
Explain it
Generative models learn patterns from examples and produce new content that fits those patterns.
Build a career on it
Families include large language models (text), diffusion models (images and video), and speech models.
ID bridge
A fast draft writer, designer, and narrator whose work always needs review.

golden set

Also called: gold standard set

A small set of examples with trusted, agreed scores, used to check raters or AI graders.

Use it
The reference review of Draft 1 works like a tiny golden set.
Explain it
New raters or AI judges are compared with the golden set to see if they apply the rubric correctly.
Build a career on it
Building golden sets is a common task in evaluation roles.
ID bridge
Like anchor papers used to train assessment graders.

grounding

Also called: grounded

Making an AI base its answer on specific sources you give it, instead of its general training.

Use it
Give Claude the extracted files and ask it to cite where each fact came from.
Explain it
Grounded answers are easier to check. A source facts table is a simple grounding tool.
Build a career on it
Retrieval-augmented generation (RAG) and citation checks are common grounding methods.
ID bridge
Like requiring every claim in a course to trace back to the SME's source.

hallucination

Also called: invented facts

When an AI states something false or made up as if it were true.

Use it
Check names, numbers, and claims against the source before you use them.
Explain it
The model produces likely text, not verified facts. It fills gaps with patterns.
Build a career on it
Teams measure hallucination rates and faithfulness to sources in evaluations.
ID bridge
Like a confident SME who misremembers a detail.

human in the loop

Also called: human review

A way of working where a person checks and approves AI output before it is used.

Use it
In this course, you decide at the image, review, revision, and publish steps.
Explain it
Human review catches errors and keeps someone accountable for the result.
Build a career on it
Designing review points is part of AI workflow and governance roles.
ID bridge
Like an SME review and sign-off step.

inter-rater reliability

Also called: agreement

How often different raters give the same score to the same work.

Use it
Count how many criteria you and another rater scored the same (exact) or within one point (adjacent).
Explain it
Simple percent agreement is quick to read; statistics like Cohen's kappa adjust for chance agreement.
Build a career on it
Evaluation teams track agreement to decide when guidelines need fixing.
ID bridge
The reliability of your assessment scoring.

invented detail

Also called: invented details

Something in AI output, in text or an image, that was not in the source.

Use it
In images, look closely at small text and numbers. If you cannot read it in the original, do not trust it in the AI version.
Explain it
Generative models add plausible detail when information is missing. It can look sharp and still be wrong.
Build a career on it
Image and text faithfulness checks are part of AI quality evaluation.
ID bridge
Like an illustrator adding labels that the engineer never approved.

job aid

Also called: quick reference card

A short reference people use while doing a task, such as a checklist or step card.

Use it
The quick reference card at the end of the guide is a job aid.
Explain it
Job aids support performance on the job; training builds understanding and skill.
Build a career on it
Performance support design transfers to AI enablement roles.
ID bridge
The same job aids you design today.

large language model (LLM)

Also called: LLM, language model

An AI system trained on huge amounts of text that writes by predicting likely next words.

Use it
It can write and edit text very well, but it can sound sure and still be wrong. Check facts that matter.
Explain it
The model learns patterns from text during training, then is tuned to follow instructions. It does not look facts up unless a tool gives it sources.
Build a career on it
Pretraining on large text collections, then supervised fine-tuning and reinforcement learning from human feedback shape the model's behavior.
ID bridge
Like a very well-read writer who has never checked a source. Great first drafts; needs an SME review.

LLM-as-judge

Also called: AI grader, AI judge

Using an AI model to score other AI outputs with a rubric.

Use it
An AI review is a second opinion. Check its evidence and do not accept its score automatically.
Explain it
AI judges are fast and consistent but can be biased, for example toward longer answers or their own style.
Build a career on it
Teams validate AI judges against human scores on a golden set before trusting them.
ID bridge
Like an automated scoring engine that still needs human moderation.

Markdown

Also called: .md

A simple text format that uses symbols like # for headings and - for lists.

Use it
The draft and the final guide are Markdown files, which are simple to review and change.
Explain it
Markdown can be turned into web pages, Word files, or PDFs with the right tools.
Build a career on it
Markdown is the common format for AI prompts, documentation, and content pipelines.
ID bridge
Like a storyboard in plain text.

Merrill's First Principles of Instruction

Also called: First Principles

A research-based design method: start from a real task, activate what learners know, demonstrate, let them apply, and help them integrate it into their work.

Use it
The starter skill uses these five principles to structure the learner guide.
Explain it
Merrill (2002) found these principles across many design models.
Build a career on it
Encoding design methods in AI instructions is a key skill for AI-enabled learning teams.
ID bridge
A method you may already use; here it becomes a skill.

PDF/UA

Also called: ISO 14289

The international standard for accessible PDF files.

Use it
A tagged PDF with a title, language, headings, and alt text is a strong start toward PDF/UA.
Explain it
Free checkers such as PAC (Windows) and veraPDF test many PDF/UA rules.
Build a career on it
Document accessibility is a specialist skill in compliance-heavy organizations.
ID bridge
Like WCAG, but for PDFs.

permission mode

Also called: Manual mode, Accept edits

The setting that decides whether Claude asks you before it edits files or runs commands.

Use it
Start in Manual so you see each change before it happens.
Explain it
Other modes let Claude accept file edits on its own, plan without editing, or use automatic safety checks.
Build a career on it
Permission design is part of AI safety and governance for agents in organizations.
ID bridge
Like approval levels in a review workflow.

personal data

Also called: PII, personally identifiable information

Information about an identifiable person, such as a name, email, phone number, or employee ID.

Use it
Remove personal data from documents before you use them with AI, unless you have clear approval.
Explain it
Privacy laws, such as the GDPR in Europe, set rules for processing personal data.
Build a career on it
Data annotation and AI training roles follow strict personal-data rules.
ID bridge
Like learner records you must protect.

plain language

Writing that the intended readers can find, understand, and use the first time they read it.

Use it
Use short sentences, common words, active voice, and define abbreviations.
Explain it
ISO 24495-1 (2023) sets plain-language principles. It helps everyone, including readers of English as a second language.
Build a career on it
AI trainer and rater roles often ask for writing, rubric, and assessment skills.
ID bridge
The writing standard you already apply to learner content.

pretraining

The first, largest stage of training, where a model learns patterns from huge amounts of text.

Use it
Pretraining is why a model knows a lot of general information, some of it out of date.
Explain it
After pretraining, models are tuned to follow instructions and be helpful and safe.
Build a career on it
Pretraining data quality and training cutoffs affect what models know.
ID bridge
Like general education before job training.

project memory (CLAUDE.md)

Also called: CLAUDE.md, memory file

A file of standing rules that Claude reads at the start of every session in a folder.

Use it
Put rules there that apply to all work in the project, and keep them short and specific.
Explain it
It is guidance, not enforcement. For hard limits, use settings or permissions.
Build a career on it
Instruction files (CLAUDE.md, AGENTS.md) are how teams share conventions with coding agents.
ID bridge
Like a project style guide that every new team member reads first.

prompt

Also called: prompts

The instruction or question you type to an AI.

Use it
Say what you want, which files to use, and what the output should look like.
Explain it
Prompts are one-time. Rules you repeat belong in project memory or a skill.
Build a career on it
Prompt design is now part of context engineering: choosing what information the model sees.
ID bridge
Like a work request or creative brief.

prompt injection

Hidden instructions inside a file or web page that try to make an AI do something you did not ask for.

Use it
Be careful with files and pages from unknown sources. Read what Claude plans to do before you approve it.
Explain it
Agents read content as input, and that content can contain text that looks like instructions.
Build a career on it
Prompt injection testing is part of AI security and red teaming.
ID bridge
Like a forged note in a stack of SME comments.

rater

Also called: raters, annotator

A person who scores AI outputs with a rubric or guidelines.

Use it
In Lesson 1.6 you act as a rater for Claude's draft.
Explain it
Raters' scores are used to check quality and to improve models.
Build a career on it
Rater, annotator, and AI trainer roles hire people with writing, assessment, and subject expertise.
ID bridge
Like a trained grader in an assessment program.

reading order

The order in which a screen reader reads the content of a page or PDF.

Use it
Check that warnings come before the steps they apply to, in both the visual and the reading order.
Explain it
In a tagged PDF, reading order follows the order of the tags, which comes from the source structure.
Build a career on it
Reading order checks are part of accessibility audits.
ID bridge
Like making sure the narration matches the order of the screen.

redaction

Also called: de-identify

Removing or hiding sensitive information from a document before sharing it.

Use it
Use the redaction checklist before you bring your own document into the workflow.
Explain it
Check text, comments, headers, file properties, and text inside images.
Build a career on it
De-identification is a routine step in AI data work.
ID bridge
Like removing learner names before sharing survey comments.

reinforcement learning from human feedback (RLHF)

Also called: RLHF

A training method where people rate or compare AI answers and the model learns to give answers people prefer.

Use it
Human ratings shape how AI assistants behave, which is why rater work matters.
Explain it
Ratings train a reward model, which is then used to adjust the AI model's behavior.
Build a career on it
RLHF and related methods create demand for raters, rubric writers, and evaluation designers.
ID bridge
Like improving a course from learner feedback, at huge scale.

rubric

Also called: rubrics

A scoring guide with criteria and descriptions of each performance level.

Use it
Score each criterion with evidence, not a general feeling.
Explain it
Analytic rubrics score criteria separately; holistic rubrics give one overall score.
Build a career on it
AI teams use rubrics to rate model outputs and to guide human raters and AI graders.
ID bridge
The same rubrics you write for assessments.

sampling

Also called: temperature

The way a model picks its next words with some randomness, so two runs give different wording.

Use it
Expect Claude's text to differ from the lesson examples. Check structure and facts, not exact wording.
Explain it
Settings like temperature control how much randomness is used.
Build a career on it
Evaluations run the same task several times to measure consistency.
ID bridge
Like two designers writing the same brief in different words.

screen reader

Also called: screen readers

Software that reads the text on a screen aloud, or sends it to a braille display, for people who are blind or have low vision.

Use it
A screen reader follows the document's tags and reading order. Check that your PDF reads in the right order and that every image has alt text.
Explain it
Screen readers use the structure of a page or PDF (headings, lists, tables, alt text) to let a person navigate and understand the content.
Build a career on it
Accessibility testing with screen readers such as NVDA and VoiceOver is a skill that product and content teams hire for.
ID bridge
The same learners you design for when you follow WCAG and write alt text.

sign-off

Also called: approval

Written approval of a document by the people who must approve it, for example the SME and the safety team.

Use it
Ask for sign-off on the exact file that you will publish, and record who approved it, in which role, and when.
Explain it
Sign-off is a human decision. An AI review can help you prepare for it, but it cannot replace it.
Build a career on it
A clear record of who approved what, and when, is part of responsible AI use in any team.
ID bridge
Like getting SME approval on a storyboard before you build the course.

skill

Also called: skills, SKILL.md

A saved, reusable set of instructions and files for one kind of job, which Claude uses when your request matches.

Use it
Put your repeatable method in a skill so you get the same structure every time.
Explain it
Only the skill's short description stays loaded until it is needed; then Claude reads the full instructions.
Build a career on it
Skills follow the open Agent Skills format, so other tools can use them too. Teams test skills with example inputs.
ID bridge
Like a documented procedure or template your team uses for every project.

SME (subject matter expert)

Also called: subject matter expert

The person who knows the content and must confirm technical facts.

Use it
When sources disagree, ask the SME. Do not let AI choose.
Explain it
AI can draft and organize, but the SME owns the facts.
Build a career on it
AI companies hire subject experts to write and check training data.
ID bridge
The same SME role you already manage in every project.

subagent

Also called: subagents

A helper agent that Claude starts for one part of a task, working in its own separate context.

Use it
Useful for doing several similar jobs at once, such as processing many documents.
Explain it
Each subagent returns a short summary, which keeps the main conversation focused.
Build a career on it
Multi-agent orchestration is a growing area in AI engineering.
ID bridge
Like delegating parts of a project to team members and collecting their summaries.

tagged PDF

Also called: PDF tags

A PDF with hidden structure (headings, lists, tables, alt text, reading order) that screen readers use.

Use it
Build PDFs from well-structured sources and check the tags before you share them.
Explain it
Tags come from the source structure. Automatic checkers find many problems, but not all.
Build a career on it
PDF/UA (ISO 14289) is the standard for accessible PDFs.
ID bridge
Like the table of contents and headings a screen reader user navigates by.

token

Also called: tokens

A small piece of text, often part of a word, that an AI model reads and writes.

Use it
Longer files and conversations use more tokens and more of your plan's limits.
Explain it
Models count limits and costs in tokens. A page of English is roughly several hundred tokens.
Build a career on it
Token counts drive pricing, speed, and context limits in AI products.
ID bridge
Like counting seat time: it is how the system measures effort.

uv

A free tool that installs Python and the right packages to run the kit's helper tools.

Use it
Claude runs commands like "uv run tools/docx_extract.py". You install uv once.
Explain it
Each tool lists the packages it needs; uv installs them in a separate space so they do not affect other programs.
Build a career on it
Reproducible tool environments are standard practice in data and AI work.
ID bridge
Like a runtime player that makes a course package work on any computer.

WCAG

Also called: Web Content Accessibility Guidelines

The international guidelines for making web content accessible to people with disabilities.

Use it
This course and its outputs aim for WCAG 2.2 Level AA.
Explain it
WCAG sets testable success criteria, such as contrast, keyboard access, and text alternatives.
Build a career on it
Accessibility audits are part of many learning and product roles.
ID bridge
The accessibility standard behind your e-learning checks.