Glossary

Terms as they are used in this course. Where the industry uses a word loosely, the definition here is the narrow one.

Acceptance criterion - A statement of behaviour precise enough to be turned into a test. Usually written as Given / When / Then. (Module 2)

Agent - A model that runs in a loop, acts through tools, observes the results and continues until a stopping rule fires. Contrast with chat and inline completion. (Modules 3 and 7)

Assistant - Any tool that generates or edits code from natural language, in any mode. (Module 1)

Audit trail - The records that answer, for any change: who merged it, which tool and model produced it, what the agent read and ran, and what review found. (Module 5)

Chunk - A piece of a document, typically 300-800 tokens, stored and retrieved as a unit in a RAG system. (Module 4)

Context - Everything a model can see when it answers: prompt, rules file, attached files, tool output. A model sees its context, not the repository. (Module 3)

Context engineering - Deliberately controlling what goes into the context: rules files, attached files, session boundaries, MCP servers. (Module 3)

Delivery loop - Intent, specification, generation, verification, release. The engineer owns four of the five steps. (Module 1)

DORA metrics - Deployment frequency, lead time for changes, change failure rate, time to restore. The baseline for measuring delivery outcomes. (Module 6)

Embedding - A vector representation of text such that similar meanings are close together. Produced by an embedding model whose version must be pinned. (Module 4)

Evaluation (eval) - A scored measurement of model-backed behaviour over a set of representative cases, compared with a threshold. Distinct from a test. (Module 4)

Evaluation gate - A CI step that fails a change when evaluation scores fall below absolute or relative thresholds. (Module 4)

Executable specification - A specification that a model can act on and a test can verify; one that a stranger could check without asking the author. (Module 2)

Fabrication - A model producing something that does not exist - an API, a flag, a fact - with full confidence. Also called hallucination. Caught by verification independent of the model. (Module 1)

Hybrid retrieval - Combining vector similarity search with keyword (BM25) search and merging the results. The standard fix for vocabulary mismatch. (Module 4)

Inline completion - The assistant mode that proposes a few lines at the cursor, accepted with Tab. (Module 3)

Intent - The three-sentence statement of what outcome is wanted, for whom, and what is out of scope. The step the assistant cannot supply. (Module 1)

Local optimum - Generated code that is correct in isolation but ignores existing helpers, conventions or middleware. Caught by checking fit with the codebase. (Module 1)

MCP (Model Context Protocol) - An open standard for connecting a model to external tools and data through servers that expose callable functions. (Module 3)

Memory - Facts a tool stores between sessions for one user. Personal; not a substitute for a shared rules file. (Module 3)

Model-as-judge - Using a second model call to grade an answer against a rubric. Must be calibrated against human grades. (Module 4)

Orchestrator - The agent (or human) that plans a task, writes contracts for workers, starts them, collects and merges their results, and reports. It does not implement. (Module 7)

Sub-agent - An agent started by an orchestrator for one bounded task, with its own context window, tools and permissions, returning a result in a defined shape. (Module 7)

Task contract - The specification scoped to one worker: context provided, done-when criteria, return shape, allowed and forbidden actions. (Module 7)

Non-goals - The one or two lines in a specification naming what is deliberately not being done. The primary defence against scope drift. (Module 2)

Prompt injection - Content in the context - a document, an issue, a web page - that acts as an instruction to the model. A design problem addressed by least privilege and human steps before consequential actions, not by filtering. (Module 5)

RAG (Retrieval-augmented generation) - Retrieve the relevant passages, give them to the model with the question, instruct it to answer only from them and cite. (Module 4)

Rules file - A committed file (CLAUDE.md, .cursor/rules/*.mdc, .github/copilot-instructions.md) that the tool reads at the start of every session; the team's constraints written once. (Module 3)

Scope drift - A diff that goes beyond the request: extra files, renames, new dependencies. Caught by comparing file count and diff size against the request and the non-goals. (Module 1)

Specification (spec) - The description of done: acceptance criteria, examples, interfaces, constraints, non-goals. Lives in the repository. (Module 2)

Test-first contract - Tests are written from the spec before the implementation, must fail for the right reason, and may not be modified by the implementation step. (Module 2)

Verification - Any check whose result does not depend on the model that produced the output: compiler, tests, linter, human review, production metric. Now the engineer's main job. (Module 1)