AI Coding Glossary

The practice of optimizing content so AI answer engines (Google AI Overview, ChatGPT, Perplexity, Claude, Gemini) cite it directly in their responses.

The core iteration cycle of an AI agent: think → call a tool → observe the result → think again, repeating until the task is done or a halt condition fires.

The layer that plans, dispatches, and supervises one or more AI agents, deciding which subagent runs when, how outputs flow between them, and when to halt.

A registry of installable AI agent skills following the SKILL.md standard, skills-hub.ai indexes 4,900+ skills from 100+ sources, with one-command install across every major AI coding tool.

The open standard that defines SKILL.md as a portable AI instruction format across every major AI coding tool.

A pattern where multiple specialized AI subagents collaborate on a task, typically implementing, testing, reviewing, and shipping in parallel.

An AI system that takes multi-step autonomous actions in a codebase, reading files, running commands, calling tools, to complete a high-level task.

Google's AI-generated summary that appears above traditional search results, synthesizing answers from multiple cited sources.

AWS's spec-driven AI IDE, write a clear specification first, then let the agent implement to spec, powered by Amazon Bedrock.

StackBlitz's browser-based AI app builder, generates and runs a full web app in WebContainer right in the browser tab.

Anthropic's terminal-native AI coding agent with sub-agents, scheduled tasks, voice mode, and first-class MCP / SKILL.md support.

An isolated execution environment where an AI agent runs code, usually a container, VM, or git worktree, to prevent agent actions from damaging the host system or main branch.

A SKILL.md that chains multiple skills into an agent team or pipeline, implement → test → review → ship as a single install.

The capability of an AI model to control a desktop or browser directly, clicking, typing, taking screenshots, to complete tasks that require GUI interaction.

The maximum number of tokens an LLM can read at once, both the prompt and the model's response combined. Larger windows allow longer documents, more conversation history, and bigger codebases.

A long-running agentic research workflow, the AI plans, searches multiple sources, synthesizes findings, and produces a structured report, distinct from quick one-shot search.

Cognition Labs' autonomous AI software engineer, a browser-based agent that operates a sandboxed VM with shell, browser, and IDE to complete multi-day engineering tasks end to end.

High-dimensional numeric representations of text, code, or images that capture semantic meaning, the input to vector search, RAG, and clustering.

Training a base LLM further on a smaller domain-specific dataset to specialize its behavior, distinct from prompt engineering and RAG.

An LLM capability where the model decides to invoke a structured function or tool with arguments, instead of (or alongside) generating text, the foundation of AI agents.

Optimizing for brand mentions and citations inside LLM-generated answers across ChatGPT, Claude, Gemini, Perplexity, and Grok, not just blue-link rankings.

Google's AI coding platform, desktop app + Go CLI + open-source SDK with a native multi-agent architecture, launched May 19, 2026 at I/O.

Rules and validation layers that constrain what an AI agent or LLM can do or say, input validation, output filtering, per-tool allowlists, approval gates, and budget limits.

When an LLM generates plausible-sounding but factually incorrect content, fabricated function signatures, invented APIs, wrong citations, or confident-but-fictional claims.

A shell command Claude Code executes automatically in response to events (PreToolUse, PostToolUse, Stop, UserPromptSubmit) configured in settings.json.

A plain-text file at a site's root that gives AI clients (ChatGPT, Claude, Perplexity) a curated map of the site's most useful pages, in a format LLMs can ingest cheaply.

AI app builder that generates full-stack web apps from a prompt, frontend, backend, database, deploy, all from the browser.

Anthropic's open standard for connecting AI clients to external tools, data sources, and capability servers via JSON-RPC.

A long-running process that exposes tools, resources, or prompts to AI clients over the Model Context Protocol.

A layer that picks the right LLM for each request, cheap model for simple tasks, frontier model for hard ones, based on prompt complexity, latency budget, or cost target.

A provider-side cache that reuses the model's processing of a repeated prompt prefix, cutting cost ~90% and first-token latency ~80% when the prefix is identical between calls.

The practice of writing inputs to LLMs to consistently get useful outputs, distinct from skills (instructions packaged as files) and fine-tuning (weight updates).

A class of attack where untrusted content hidden in an LLM's input (a webpage, a document, a tool result) overrides the system prompt and makes the model execute the attacker's instructions.

An LLM pattern where the model retrieves relevant documents from an index at query time and uses them to ground its answer, reducing hallucination and adding fresh data.

Replit's browser-based AI agent that builds and deploys full applications from a natural-language prompt, code, database, hosting, all in the Replit cloud.

Always-on, tool-specific project context files (.cursorrules, .clinerules, .windsurfrules) that load on every conversation, distinct from on-demand skills.

Claude Code's cron-style background agent runs, execute skills on a schedule without keeping a terminal open.

Structured data embedded in a page's HTML that tells search engines and AI clients what the content represents, Article, FAQ, Product, Recipe, Person, Organization, etc.

A reusable, versioned, natural-language instruction set for AI coding assistants, packaged as a SKILL.md file with YAML frontmatter.

A public registry of AI agent skills you can install in one command, skills-hub.ai is the largest with 4,900+ skills from 100+ sources.

The file format of the open Agent Skills standard, Markdown with YAML frontmatter that defines a portable AI instruction set.

A user-typed `/name` in an AI coding tool that triggers a registered skill, prompt, or built-in action.

An AI coding pattern where you write a clear specification first, then let the agent implement to spec, the inverse of vibe-coding.

An isolated, parallel AI agent spawned by a parent agent for a specific specialized task, each with its own context window and tool access.

The benchmark suite for measuring AI agents on real-world software engineering tasks, fix this GitHub issue, given the repo and tests. Top variants: SWE-bench Verified (humans confirmed solvability), SWE-bench Lite, SWE-bench Multimodal.

Search that ranks results by semantic similarity using embeddings, finding documents that mean the same thing as the query, not just documents that contain the same words.

An AI coding style where you describe what you want in chat and let the AI improvise, the opposite of spec-driven development.

Push-to-talk voice input in AI coding tools, Claude Code's /voice (March 2026) supports 20 languages.