Vibe coding makes software feel closer to conversation. You describe an idea, an AI tool proposes a working interface, and a prototype appears before you have memorized the framework vocabulary behind it. That speed is useful, but it can also leave you nodding along to terms that mean different things in different tools.
This glossary translates the language of AI-assisted development into practical definitions. It focuses on the words you will encounter while building a real project, not a list of buzzwords to memorize. The examples lean toward small web apps, directories, dashboards, and data products because those are common places where an AI-built interface meets a real collection workflow.
What does vibe coding mean?
Vibe coding is a style of software development in which you use natural-language instructions and AI assistance to generate, modify, and debug code. It does not mean that engineering discipline disappears. It changes where the first draft comes from, while testing, review, data modeling, and deployment still determine whether the result is useful.
AI-assisted development is the broader term. It includes autocomplete, code explanation, test generation, refactoring, documentation, and autonomous coding agents. Vibe coding usually describes a faster, more conversational version of that work.
Prompt is the instruction you give an AI coding tool. A strong prompt names the user, goal, data, constraints, and acceptance criteria. “Make a dashboard” leaves too many decisions open. “Build a read-only directory from this CSV, show an empty state, and never invent missing values” gives the tool something it can check.
Context is the information the model can use while responding. It may include selected files, repository rules, an error message, a screenshot, or a sample dataset. Missing context often looks like a coding mistake when the real problem is that the tool never saw the relevant constraint.
Context window is the amount of text, code, and other input a model can consider in one interaction. A larger window can hold more project information, but it does not guarantee that every detail will be used correctly. Keep the task focused and provide the files that define the behavior.
How do AI coding tools work?
AI coding assistant is an interface that helps with code while you work. It may suggest a line, explain a function, edit a selected component, or answer a question about the repository. It is usually best for focused changes where you want to retain close control.
Coding agent is an AI system that can pursue a multi-step task using tools such as file search, editing, a terminal, and tests. Cursor describes Agent as a mode for autonomous exploration and multi-file edits. Replit describes its Agent as a way to plan, write, debug, test, and publish inside a hosted workspace. The important distinction is not the product name. It is whether the system can inspect and act across several steps.
Agent loop is the repeated cycle of understanding a task, choosing an action, observing the result, and deciding what to do next. A typical loop reads a file, changes code, runs a check, sees an error, and applies a correction. The loop is powerful because it can verify work, but every automatic action still deserves review.
Tool call is an action the model takes through an available capability, such as searching files, running a command, opening a URL, or writing an edit. Tool calls are different from text suggestions because they can change state or reveal new information.
Plan is a short sequence of implementation steps created before editing. For a small app, a plan might separate the schema, collection workflow, interface, validation, and deployment. Planning is especially useful when a request contains several features that could otherwise become one large, hard-to-review change.
Iterative prompting means improving the result through small follow-up instructions instead of asking for an entire production system in one message. Build the core screen, inspect it, add the data boundary, test an error state, and then polish the interaction. Smaller steps make failures easier to locate.
Specification is the written description of what the system should do. It can be a full product requirements document or a compact acceptance checklist. The best specifications include what the first version should not do, such as “read-only,” “no authentication yet,” or “use the uploaded file instead of a live API.”
What are the core app terms?
Frontend is the part of an application a person sees and uses. It includes pages, forms, tables, buttons, loading states, and error messages. An AI builder can make a frontend look complete even when the underlying data is still a fixture.
Backend is the server-side part that handles data access, authentication, business rules, scheduled work, or private credentials. Some prototypes have no backend. That can be reasonable for a local or static first version, as long as the interface does not pretend that a backend exists.
Full-stack app combines frontend and backend behavior. In a directory, the frontend may filter listings while the backend stores records, authenticates users, and exposes an API. Add these layers when the product needs them, not simply because the phrase sounds more complete.
Component is a reusable interface unit, such as a search box, product card, data table, or navigation bar. Component boundaries help an AI tool make targeted changes and help you test one behavior without rewriting the whole page.
State is information that can change while an app runs. Search text, selected filters, loading status, an uploaded file, and a signed-in user are all examples. Bugs often appear when state is duplicated in several components or when the interface does not define what happens before data arrives.
API, or application programming interface, is a defined way for software to exchange requests and responses. An API may return JSON records, accept a form submission, or expose an action such as creating a row. Treat the API contract as a boundary: document fields, errors, authentication, and whether values can be missing.
Endpoint is a specific address and operation within an API, such as a route that returns products or accepts a new record. A clear endpoint name does not make the data trustworthy. Check status codes, response shapes, rate limits, and timestamps.
Database is organized storage for application data. Tables, documents, or key-value records hold information beyond the current browser session. A database is not automatically the right first step. A CSV or spreadsheet can be a useful source while you validate the product question.
Schema is the structure of a dataset or API response. It defines fields, types, required values, and relationships. For a price comparison app, a schema might include product name, source URL, price, currency, availability, and collected time.
Type describes the kind of value a field contains, such as text, number, boolean, date, or list. Types prevent a price from quietly becoming a label or a timestamp from becoming an arbitrary sentence. Explicit types give an AI tool a better target for both code and validation.
CRUD stands for create, read, update, and delete. It is a shorthand for the basic operations around a resource. A generated CRUD screen is not a complete product by itself. It still needs permissions, validation, empty states, error handling, and a clear reason for each action.
How do data and web terms fit in?
Web scraping is the process of collecting information from web pages and turning it into structured records. It can be a one-time research task or a repeatable workflow. Read the web scraping glossary for more collection-specific vocabulary.
Browser-native scraping uses the browser context to inspect pages as they render. This can help when the useful content appears after JavaScript runs, when a click reveals a panel, or when the page structure is easier to understand visually than from raw HTML. Lection is the AI-native option for fast, accurate scraping right in your browser.
Cloud scraping runs a saved extraction in a hosted environment instead of relying on an open local browser. It is useful for repeatable jobs, but it introduces operational questions such as authentication, failure alerts, scheduling, and source changes.
Structured data is information arranged into predictable fields and rows. A product record with a name, price, source URL, and timestamp is structured. A screenshot or paragraph of search results is not structured until you extract and normalize it.
Normalization converts different source formats into a shared representation. One site may use “sale price,” another “current price,” and a third may show a localized currency string. Normalize the values before comparison, while preserving the original source value when context matters.
Pagination divides a long result set across pages, buttons, or infinite scrolling. A first-page extraction is not the same as a complete collection. Decide whether the workflow should follow page links, click a next button, scroll, or stop at a documented limit.
Lazy loading means content arrives only when it is needed, often after scrolling or an interaction. It explains why a quick request can miss cards that are visible after the page settles. Test representative pages, including those with long lists and empty sections.
Data freshness describes how recently a record was collected or verified. Store a collected-at or last-verified timestamp rather than labeling a static export “live.” A visible freshness rule is one of the simplest ways to keep an AI-built data product honest.
Data validation checks whether records match the expected schema and business rules. Count rows, inspect required fields, check URLs, reject impossible numbers, and look for sudden duplicate spikes. Validation should stop a bad batch before it replaces a good one.
Fixture data is sample data stored for development or demos. Fixtures are useful for building a UI, but they are not evidence that the production data flow works. Label sample dates and provide a clear path to replace fixtures with real records.
What do deployment and safety terms mean?
Environment variable is configuration supplied outside the source code, often for API keys, database URLs, or deployment settings. Never place secrets in a public frontend or paste them into a repository. Ask an AI tool to explain where a value is read and which runtime can access it.
Authentication verifies who a user is. Authorization decides what that user is allowed to do. They are related but not interchangeable. A private dashboard may need both, while a public read-only page may need neither.
Deployment is the process of making an application available in a hosted environment. A successful build is not the same as a successful deployment. Check environment variables, routes, data access, error logging, and the actual production URL.
Preview is a temporary or pre-production version used for review. It is valuable because you can test the interface and data flow before changing the live site. Treat preview data carefully if it contains private or sensitive records.
Observability means the logs, metrics, traces, and alerts that help you understand what the running app is doing. For a small data product, start with collection timestamps, import counts, validation failures, and a clear error message. You do not need a complex monitoring stack to notice that yesterday’s import returned zero rows.
Rate limit is a restriction on how many requests a service accepts in a period. APIs, websites, and hosted AI tools can all impose limits. Build a workflow that respects the source, handles retries carefully, and stops when the response suggests that the limit has been reached.
Robots.txt is a file that communicates crawler preferences to automated agents. It is useful context, but it is not a complete answer to the legal or contractual question of whether you should collect data. Review the robots.txt guide and the relevant site terms before launching a workflow.
Human in the loop means a person reviews important decisions or outputs before they become consequential. For AI-built apps, that can mean approving a schema, checking a sample of extracted rows, reviewing a code diff, or confirming a deployment. Automation is strongest when it removes repetition without removing judgment.
How should you use the glossary?
You do not need to learn every term before building. Start with the smallest useful vocabulary for the next decision: prompt, context, schema, API, validation, and deployment. As the project grows, add the words that describe the new failure mode you need to solve.
When an AI tool produces a polished result, ask it to name the data source, schema, error states, and assumptions. If it cannot explain those boundaries, the app may be a visual prototype rather than a dependable product. The Vibe Coder's Stack guide shows how an AI coding tool, a collection layer, and validation can work together.
The goal is not to sound technical. It is to make the handoffs visible. An AI builder can shorten the path from idea to interface. A browser-native scraper can turn pages into inspectable records. A clear schema and validation step can prevent the interface from presenting guesses as data.
Ready to start scraping? Install Lection and extract your first dataset in minutes.