The Vibe CoderGlossary

The repeated cycle an AI agent runs — perceive → plan → act → observe — until a goal is reached or it's stopped. The engine inside any agentic workflow.

A multi-step AI process where the model takes autonomous actions toward a defined goal — calling tools, making decisions, and self-correcting — with minimal human intervention.

Application Programming Interface. The connection layer between your app and an external service. When you call Claude, Calendly, or FUB from a workflow, you're using their API.

A discrete, usable output produced by an AI — a document, code file, image, structured JSON, or report. Artifacts are what you hand off, deploy, or send to clients.

A system prompt component that defines the AI's tone, persona, style, and boundaries. Converts a generic LLM into a specialized communication engine tied to your brand.

A structured prompt document handed to an AI coding agent (Manus, Cursor, Lovable) that defines what to build, the stack, constraints, and expected outputs. Replaces a traditional PRD.

A prompting technique that asks the model to reason step-by-step before producing a final answer. Forces deliberation over pattern-matching. Measurably improves accuracy on complex tasks.

When a long conversation causes the AI to lose track of earlier instructions, facts, or tone — producing outputs that contradict or ignore what was established upfront.

The total amount of text (input + output) an LLM can hold in working memory for a single session. Everything inside it is 'seen.' Everything outside it doesn't exist.

A prompt that consistently fails to produce usable output regardless of small tweaks. Requires a full rebuild from a different angle — not iteration.

Where your AI-powered system runs for real users — a web app, an embedded widget, a triggered workflow, or an API endpoint. The bridge between your build and your audience.

A hidden configuration value — API keys, system prompts, credentials — stored outside the codebase and injected at runtime. Keeps sensitive data out of your source code.

Training a base LLM on a custom dataset to specialize its behavior, vocabulary, or style. Produces a model variant that defaults to your patterns without extensive prompting.

Starting a new conversation branch in an AI session to test an alternative approach without losing the original thread. Low-cost way to run parallel experiments.

A model capability that lets an LLM trigger external functions, APIs, or tools mid-response — instead of just generating text, it takes action. The core mechanic of agentic systems.

Anchoring AI output in specific real data — documents, retrieved facts, client records — to reduce hallucination and increase relevance. The practical outcome of RAG.

When an AI confidently generates false, fabricated, or contradictory content — presenting invention as fact. A structural property of all current LLMs, not a bug.

The computational act of running a trained model to generate output. Every time you prompt Claude, you're triggering inference. Training ≠ inference — training is past tense.

The vibe coder's core development cycle: prompt → output → evaluate → refine → repeat. Replaces traditional write-compile-debug cycles. Speed of iteration is your competitive edge.

The delay between sending a prompt and receiving a model response. Critical for real-time user-facing features. Less critical for backend batch processing.

Large Language Model. The neural network foundation underneath tools like Claude, GPT, and Gemini. Trained on massive text datasets to predict and generate language — and, with prompting, reason.

Deliberately directing different tasks to different AI models based on complexity, cost, and required capability. A cost and quality optimization layer for production systems.

Multiple AI agents working in coordination, each with a specialized role — one plans, one executes, one validates. Enables tasks too complex or long for a single agent.

Building functional, production-grade AI-powered systems without writing traditional code — using natural language prompts, visual builders, and API connections instead of syntax.

Coordinating multiple AI agents, tools, and data sources to execute a complex workflow. The conductor role — managing who does what, when, and in what order.

Constraining or formatting AI output for downstream use — specifying JSON, markdown, numbered lists, or structured fields so the response plugs directly into the next system.

Stored context that survives across AI sessions — user profiles, past decisions, conversation history. Requires a database layer. Without it, every session starts from zero.

The craft of writing instructions that reliably extract high-quality, specific outputs from an LLM. Covers framing, structure, role assignment, constraint setting, and output formatting.

An attack where external input — user text, imported data, scraped content — hijacks or overrides an AI system's intended behavior by containing conflicting instructions.

Retrieval Augmented Generation. Feeding external documents or database records into the LLM's context at query time, so it reasons over your data without being retrained.

The surrounding structure — UI, triggers, data pipelines, authentication — that makes an AI model useful in a real product. The model is the brain; scaffolding is the body.

Temporary data held during a user's active session that disappears when they leave. Not memory — a scratchpad. Useful for multi-turn conversations, useless for cross-session continuity.

The foundational, hidden instruction set that governs an AI agent's persona, behavior, boundaries, and output style. Runs before any user input. Defines who the AI is in your system.

A model parameter controlling output randomness. Low (0–0.3): precise, consistent, predictable. High (0.7–1.0): creative, varied, surprising. Tune per use case.

The fundamental unit of LLM processing — roughly 3/4 of a word in English. Context windows, pricing, and output limits are all measured in tokens.

Giving an LLM the ability to trigger external functions — web search, database queries, API calls, code execution — mid-response. The core capability that separates agents from chatbots.

A rapid qualitative evaluation of an AI output's overall tone, accuracy, and fit before committing to refinement. Not a detailed audit — a gut-level pass/fail on direction.

A builder methodology where natural language, systems thinking, and outcome description replace syntax as the primary development interface. The AI handles implementation; the human handles intent and judgment.

An automated HTTP trigger that fires when an event occurs — a form submission, a payment, a status change — passing data to another system in real time. The connective tissue of automation stacks.

A product built on top of an existing AI model's API, adding custom UX, system prompting, and data layers to create a specialized experience. Most AI SaaS products are wrapper apps.