Glossary

The specific, testable conditions that must be true for a piece of work to count as finished. They turn a vague request into a checklist everyone agrees on before work starts.

A short document that captures one significant technical decision, the context behind it, and its consequences. ADRs live next to the code so the reasoning survives team changes.

An AI system that takes actions on its own — calling tools, reading files, running commands — to reach a goal, instead of only returning text. It loops: observe, decide, act, repeat.

A defined way for two pieces of software to talk to each other. It specifies what requests you can make and what you get back, without exposing the internal workings.

A document describing the technical requirements, constraints, and quality attributes a system must satisfy — the architectural counterpart to a PRD's product focus.

The ordered list of everything that might be worked on — features, fixes, ideas — that hasn't been started yet. The top is prioritized; the bottom is 'someday maybe.'

Standard, repetitive code or text that must be present but carries little unique meaning — the setup you copy into every new file or project.

A prompting technique where the model is asked to reason step by step before giving a final answer. Showing its work often improves accuracy on complex tasks.

Also known as: CoT

An automated pipeline that tests every code change and ships it with minimal manual steps. CI checks the change merges cleanly and passes tests; CD deploys it.

The maximum amount of text — your prompt plus the model's reply — an AI can consider at once, measured in tokens. Exceed it and earlier content gets dropped.

A shared, explicit standard for when work is truly complete — not just 'code written' but tested, reviewed, documented, and deployable. It keeps 'done' from meaning different things to different people.

Producing the same output every time for the same input. Most code is deterministic; LLMs are not by default, which is why the same prompt can give different answers.

An unusual or extreme situation at the boundary of normal use — an empty list, a huge file, a leap year. Software breaks at edge cases far more than on the common path.

A list of numbers that represents the meaning of a piece of text so a computer can compare it to other text. Similar meanings produce similar embeddings.

A single addressable URL on a server that performs one operation — 'create an invoice' or 'fetch a user.' An API is made up of many endpoints.

A large body of work too big for one unit, broken into smaller stories or issues. An epic groups related work toward a single bigger outcome.

Giving the model a handful of examples of the task inside the prompt so it can mimic the pattern. Contrasts with zero-shot, where you provide none.

Also known as: Few-shot

Further training a base model on your own examples so it specializes in your tone, format, or domain. More involved than prompting, but more consistent for narrow tasks.

A mechanism that lets a model request a specific tool with structured arguments, so it can fetch data or take actions instead of only writing text.

Also known as: Tool Use

Tying an AI's answers to real, verifiable source material — your documents, a database, search results — so it relies on facts rather than its own guesses.

Rules and checks that constrain what an AI can say or do — blocking unsafe output, enforcing a format, or limiting which tools it may use.

When an AI states something false or invented as if it were fact. It happens because models predict plausible text, not verified truth.

An operation you can run repeatedly with the same effect as running it once. 'Set status to paid' is idempotent; 'add $10' is not.

The act of running a trained model to get a result — actually using the AI, as opposed to training it. Each request you send is one inference.

A single tracked unit of work — a bug, feature, task, or question — with its own description, status, and discussion. The atomic unit of 'what needs doing.'

A simple, human-readable text format for structured data, used almost everywhere to pass information between systems — and to get reliable structured output from AI.

A structured collection of an organization's documents and answers that an AI can draw on to respond accurately. Often the source feeding a RAG system.

An AI trained on vast amounts of text to predict and generate language. It powers chat assistants, coding tools, and writing aids.

An open standard for connecting AI assistants to external tools and data through a common interface, so the same tool works across different AI apps.

A stand-in for a real component used during testing. It imitates the real thing's behavior so you can test code in isolation without the real dependency.

Also known as: Stub

The smallest version of a product that delivers real value and can be put in front of users to learn from. It tests the idea before heavy investment.

The field of AI focused on enabling computers to understand and generate human language. Modern LLMs are its most visible result.

Giving the model exactly one example of the task before asking it to perform — a middle ground between zero-shot and few-shot.

A document that defines what a product or feature should do, for whom, and why — the problem, the users, the requirements, and how success is measured.

The instructions and context you give an AI to get a result. The quality and structure of the prompt strongly shape the quality of the output.

An attack where hidden or malicious instructions in content the AI reads cause it to ignore its real instructions — a key risk when an AI processes untrusted text.

A technique where the system first retrieves relevant documents, then feeds them to the AI so its answer is grounded in your specific data rather than general training.

Restructuring existing code to make it cleaner or easier to change without altering what it does. Behavior stays the same; the internals improve.

A bug that breaks something that used to work, usually introduced by a recent change. Automated tests guard against regressions.

A written proposal shared for feedback before a decision is locked in. It lays out a problem and a suggested approach so others can weigh in early.

An automatically generated starting structure for new code — the folders, files, and boilerplate — so you begin from a working skeleton instead of a blank page.

A formal description of the shape data must take — its fields, types, and rules. Schemas validate input and let an AI return predictable structured output.

The agreed boundary of what a project or task includes — and, just as importantly, excludes. 'Scope creep' is when that boundary quietly expands.

A place in the code where you can change behavior without editing in that exact spot — a natural boundary where parts connect. Good seams make code easy to test, extend, and hand to an AI to modify safely.

A precise description of how something should behave or be built. More detailed than a PRD's 'what,' a spec often covers the exact 'how.'

A fixed, short time-box — often one or two weeks — in which a team commits to completing a set of work. It creates a regular rhythm of planning and delivery.

The behind-the-scenes instructions that set an AI's role, rules, and tone for an entire conversation, separate from what the user types each turn.

The future cost of choosing a quick solution now over a better one. Like financial debt, it accrues 'interest' — slower changes and more bugs — until paid down.

A setting that controls how random an AI's output is. Low temperature gives focused, predictable answers; high temperature gives more varied, creative ones.

The unit AI models read and write in — roughly a word-piece, about three-quarters of a word in English. Context limits and pricing are measured in tokens.

A thin, end-to-end slice of a feature built first to prove the whole path works — from interface to database — before fleshing out the details.

A short description of a feature from the user's perspective: 'As a [role], I want [goal] so that [benefit].' It keeps work focused on user value.

A database built to store embeddings and find the most similar ones fast. It's the retrieval engine behind most RAG systems.

A complete, working piece of functionality that cuts through every layer — interface, logic, and data — for one small capability, rather than building one layer at a time.

An automated message a service sends to a URL when an event happens, so systems react in real time instead of repeatedly asking 'anything new yet?'

Asking a model to do a task with no examples, relying only on its instructions and prior training. Contrasts with few-shot.

Also known as: Zero-shot