Glossary

Each term below has its own stable link that other pages can use. A definition describes how this guide uses the word, and “the framework” means the five-stage method this guide describes.

Agent

An AI system that works toward a goal by taking steps and using tools, such as reading files or running commands.

Agentic platform

Software in which an agent can use tools, read and write files and run steps.

Alignment matrix

A grid that shows how well each chapter prepares learners for each exam skill, rated high, medium, low or none. The certification alignment workstream produces it.

Audit trail

Layer 3 of the verification layers. The framework specifies it as a record of every check made on the content, every problem found and every correction. Expert review is a separate layer, Layer 2.

Batch

A small group of units of work, such as sources, that are processed together in one run. Work can pause after each batch at a gate for a person to say whether to continue.

Blueprint

The program’s own plan, which groups learning objectives (what a learner should be able to do) into weighted content domains (major topic areas). A certifying body’s published list is a certification outline, not a blueprint.

Capability

A feature that a platform must offer for a prompt to work, such as reading files, running commands in a terminal or searching the web. Each prompt names what it needs from a fixed list of ten, given in the authoring conventions. Platform requirements says how to supply each one.

Certification alignment

A parallel workstream, not a stage, that rates each chapter against each exam skill in a certification outline. The ratings form an alignment matrix and are meant to show learners where to spend study time.

Certification outline

The list of skills that a certifying body (the organization that awards a certification) publishes for an exam. Each entry on it is an exam skill. The program’s own plan is the blueprint, not this list.

Condensation

Shortening text by applying named techniques, instead of cutting it as you go. In the framework, AI makes condensation passes over drafted text.

Context window

The amount of text an AI model can take in at one time. When a long task fills it, work can continue in a fresh session with a hand-off document.

Distractor

A plausible wrong answer option in a multiple-choice test question. In the framework, Stage 5 extracts its own misconceptions afresh from each chapter’s own text and builds distractors from those, not from the misconceptions recorded in Stage 1. The question part is the stem.

Dry run

A script mode that shows what the script would write or delete, without doing it. By this guide’s convention, a script that writes or deletes files runs as a dry run unless you pass --write. The mock model provider, which lets a script that calls an AI model run offline, is a different thing.

Exam skill

One entry in a certification outline: one skill that the exam covers. Its ID has the form CB-d.n, as in CB-3.1.

Gate

A checkpoint where a person approves before work continues. The project notes also use “gate” for automated script checks; this guide means a person’s approval unless it says otherwise.

Grounding

Making an AI answer from text it is given, with citations, instead of from what the model learned in training. In the framework, the AI tutor is designed to answer from course material in this way.

Hallucination

Output from an AI model that sounds confident but is false or invented. It is an error in what the AI wrote, unlike a misconception, which is a wrong belief a person holds.

Hand-off document

A file that lets a fresh session, meaning a new working conversation with an agent, continue a task. It records the goal, the current task and the next action. The framework specifies keeping it current (in the reference implementation, after every batch) and starting a new session when the context window runs low.

Human in the loop

A design in which a person makes some of the decisions and approves work at set points, called gates in this guide. The Human roles, gates and batching page lists which decisions the framework assigns to people.

Integrity guardrail

In the framework, a tutor rule for graded work. Instructors decide whether AI help is allowed on graded work; when it is not, the tutor’s integrity guardrail withholds full solutions and offers hints. The framework specifies that the tutor applies it before choosing a coaching protocol.

Item bank

A tagged pool of test questions organized against a blueprint; sources also call it a question bank.

Job-task analysis

A list of the duties and tasks of a job, with the knowledge, skills and abilities needed to do them.

Knowledge base

The whole collection of knowledge items extracted from a program’s sources, arranged in a prerequisite hierarchy.

Knowledge item

An extracted content record: a small self-contained idea in your own words, with a type such as definition or mechanism, links to related items, the source passage it came from (see provenance) and a review status. Together the items form the knowledge base. Neither a test question nor an exam skill is a knowledge item.

Misconception

A specific wrong belief that many people hold about a topic. In the framework, misconceptions are recorded for each knowledge item in Stage 1. Stage 5 records a further, separate set of misconceptions afresh from each chapter’s own text, and builds distractors from those, not from the Stage 1 records.

Misconception catalog

In the framework, a numbered list of the misconceptions for one chapter. The tutor uses it to diagnose learner errors. It is not used to write distractors, which Stage 5 builds from its own misconceptions, extracted afresh from each chapter’s own text, not from the catalog and not from the Stage 1 records.

Orchestrator

An agent that coordinates helper agents, called subagents. In the reference implementation, its written rules (the project notes call them the contract) give it the accept and merge decisions. It is an AI session, not a person.

Placeholder

A named gap in a prompt, such as OBJECTIVE_ID, that you replace with a real value. It is written in capital letters with underscores, inside doubled curly braces.

Prerequisite hierarchy

Content sorted into four levels, from foundational to applied, where each level builds on the levels below it.

Prompt

A written instruction that you give to an AI model. In this guide each prompt is one file, and it may contain placeholders that you fill in.

Prompt injection

An instruction hidden in a document or web page and aimed at the AI that reads it. Treat text an AI fetches as data, never as instructions.

Protocol (tutor)

In the framework, a coaching protocol: a named, reusable teaching pattern for the AI tutor, with a trigger that says when to use it and a script of steps. It is a design for teaching, not the prompt text that tells the AI what to do.

Provenance

A record of where a knowledge item came from: the source document, the section and, where possible, the page.

Rate limit

A cap that a service puts on how many requests you may make in a period.

Reference implementation

The original program and tooling this guide was written from. Where the guide gives numbers from it, they are that project’s parameters, not universal rules. Its own working documents are called the project notes in this guide.

Stage

A major part of the framework. The framework has five stages that run in order and together build an upskilling program. They are described in the Pipeline overview.

Stem

The question part of a multiple-choice test question, written to make sense without its answer options.

Sub-stage

A numbered part of a stage, with an ID such as S1.4a, where S1 is the stage.

Subagent

A helper agent that another agent starts for one part of a task. Several can run at the same time.

Upskilling program

A structured course of instruction, practice and assessment that aims to bring working professionals to competency in a topic, measured against an external standard where one exists.

Verification layer

In the framework, one of three layers of checking that the framework specifies for drafted content. Layer 1 is automated checks that look for errors such as hallucinations. Layer 2 is expert review, in which a subject-matter expert checks the content. Layer 3 is the audit trail, a record of the checks.

Easily confused pairs

Each row compares terms that are easy to mix up.

Pair	The difference
Knowledge item vs knowledge base	A knowledge item is one small idea. The knowledge base is the whole collection of them.
Blueprint vs certification outline	A blueprint is the program’s own plan. A certification outline is a certifying body’s list of exam skills.
Certification outline vs exam skill	The outline is the whole published list. An exam skill is one entry on it.
Misconception vs distractor	A misconception is a wrong belief. A distractor is a wrong answer option built from one.
Misconception catalog vs distractor	The catalog is the tutor’s per-chapter list for diagnosing learner errors. Distractors are built from misconceptions Stage 5 extracts afresh from each chapter’s own text, not from the catalog and not from the Stage 1 records.
Hallucination vs misconception	A hallucination is false or invented output from an AI model. A misconception is a wrong belief a learner holds.
Protocol (tutor) vs prompt	A protocol is a teaching pattern. A prompt is the text that instructs the AI.
Agent vs subagent vs orchestrator	An agent is an AI that works toward a goal. A subagent is a helper agent for one part of a task. An orchestrator coordinates subagents and, in the reference implementation, is given the accept and merge decisions.
Expert review vs audit trail	Expert review is Layer 2 of the verification layers: people check the content. The audit trail is Layer 3: the record of checks, problems and corrections.
Gate vs an automated script check	A gate is where a person approves before work continues. An automated script check is not a gate in this guide, although the project notes sometimes call it one.
Dry run vs the mock model provider	A dry run is about writing files: it shows what would be written. The mock model provider is about calling an AI model: it lets a script run offline.