An AI model trained on massive amounts of text to understand and generate natural language. Claude, GPT-5, and Gemini are LLMs: it predicts the next word with such quality that the result is useful reasoning, writing, and code. It is used via chat, API, or built into tools.
Token
The smallest unit of text with which an LLM processes language: it can be a word, part of a word, or a symbol. Models charge per million input and output tokens, and their context window is measured in tokens. In English, 1 token equals on average about 4 characters.
Context window
The maximum number of tokens a model can 'see' at once in a conversation or task, including your instructions, files, and its responses. A large window makes it possible to work with entire books or whole codebases; Claude and 2026's high-end models exceed 200,000 tokens.
Hallucination
An AI model response that sounds confident but is incorrect or invented: citations that don't exist, APIs that never existed, fabricated data. It is reduced with clear prompting (asking for cited sources), RAG to provide real data, and human verification for high-impact decisions.
Multimodal
A model's ability to process and combine several data types: text, images, audio, and video. A multimodal model can describe a screenshot, analyze a chart, or transcribe audio and reason about it in the same conversation.
Reasoning model
A model that dedicates an internal 'thinking' phase before responding: it breaks down the problem, evaluates paths, and checks its logic. Examples: Claude with extended thinking, OpenAI's o3. It costs more tokens and time, but improves at math, complex code, and multi-step analysis.
Generative Engine Optimization (GEO)
Content optimization so that generative engines (ChatGPT, Perplexity, Claude, Gemini) cite it in their responses. Key techniques: a direct answer in the first 60 words, structured data (Schema.org), llms.txt, and citable original statistics. It is to SEO what AI answers are to the blue list of links.
Embedding
A numerical representation (vector) of a text's meaning that lets machines compare semantic similarity. Two texts with similar meaning have nearby vectors: it is the basis of semantic search, vector databases, and RAG.
Retrieval-Augmented Generation (RAG)
A technique that combines search and generation: before responding, the system retrieves relevant documents (with embeddings) and injects them into the model's context. Result: responses anchored to your real data, with fewer hallucinations and the ability to cite the source.
Vector database
A database specialized in storing and searching embeddings by semantic similarity instead of exact match. Qdrant, Pinecone, pgvector, or Chroma enable 'search by meaning' and are the typical store for the R in RAG.
Fine-tuning
Additional retraining of a base model with your own data to specialize it in a domain, tone, or format. Nowadays, for most cases it is more cost-effective to use advanced prompting and RAG; fine-tuning is reserved for very stable format needs or highly specialized domains.
Open source
Software whose code is public and can be used, modified, and redistributed under a license (MIT, Apache 2.0...). The resources curated in PackIA (skills, MCP, agents, prompts) are open-source GitHub resources: the added value lies in curation, documentation, and organization by niche.
Agents and MCP
AI agent
An AI system that executes complete tasks autonomously: it receives a goal, plans steps, uses tools (search, write code, call APIs), and verifies its own result. Unlike a chat, it doesn't just respond: it acts. Claude Code and Codex CLI are development agents.
Subagent
A specialized agent that lives inside another orchestrator agent: it has its own system prompt, allowed tools, and a description of when to activate. In Claude Code they are defined as markdown files in .claude/agents, and the main agent delegates tasks to them to keep context clean.
Model Context Protocol (MCP)
An open protocol (created by Anthropic) that standardizes how AI models connect to external tools and data: files, GitHub, Notion, databases, browsers. An 'MCP server' exposes tools that the model invokes autonomously. It is the 'USB-C' of the AI ecosystem.
MCP server
A program that implements the MCP protocol and exposes a set of tools for a specific service (for example: GitHub's MCP exposes creating issues and PRs; Postgres's, querying the database). It is configured in the client (Claude Desktop, Claude Code, Codex CLI) and stays available in every session.
Multi-agent orchestration
Coordination of several specialized agents collaborating on a task: one researches, another writes, another reviews. Frameworks like CrewAI, AutoGen, and LangGraph implement orchestration patterns (sequential, debate, hierarchical). More power, but also higher API cost and more points of failure.
Agentic loop
An agent's work cycle: observe the state → think → act with a tool → observe the result → repeat until the goal is complete. It is what lets Claude Code edit a file, run tests, read the error, and fix it without human intervention at every step.
Tool use / function calling
An LLM's ability to invoke developer-defined functions when the task requires it: searching the web, querying an API, running code. The model decides when and with what arguments to call each function; it is the underlying mechanism used by agents and MCP servers.
Claude Code
Anthropic's tool that turns Claude into a development agent in your terminal: it reads your code, edits files, runs commands and tests, and makes commits. It works with CLAUDE.md as the project's memory, supports subagents, skills, and MCP servers. It is Codex CLI's direct competitor.
OpenAI Codex CLI
OpenAI's terminal development agent, the equivalent of Claude Code: it operates on your repository with GPT models, runs commands, and edits code. It is configured with config.toml and supports MCP servers. Many professional workflows keep both to compare results.
CLAUDE.md
An instructions file in a project's root that Claude Code reads automatically on startup: code conventions, build and test commands, architecture, and team rules. It is the agent's 'permanent memory' in your repository and the standard way to standardize its behavior.
Rate limit
The limit on requests or tokens an AI provider allows per minute/day under your plan. Intensive agents hit them fast: it pays to choose a model per task (mini for routine work), cache results, and watch the API usage headers.
Prompts
Prompt
An instruction you give an AI model: a question, a task, or a combination of context, role, and desired output format. Result quality depends directly on prompt specificity: an expert role, concrete context, a defined deliverable, and clear constraints produce professional responses.
Prompt engineering
The discipline of designing, testing, and systematizprompt engineering guide to get consistent results from an LLM. It includes techniques like few-shot (giving examples), chain-of-thought (asking for step-by-step reasoning), section delimiting, and templates with variables. In production it is versioned like code.
System prompt
Background instruction that defines the model's permanent behavior before any conversation: identity, tone, limits, and format. In a Claude Code subagent, the system prompt IS its definition; in AI products, it is where your brand personality lives.
Few-shot prompting
A technique of including desired input-output examples inside the prompt so the model imitates the pattern: two or three well-chosen examples are usually worth more than a paragraph of instructions. The opposite extreme is zero-shot: only instructions, no examples.
Chain-of-thought (CoT)
A prompting technique that asks the model to reason step by step before the final answer, improving accuracy in math, logic, and analysis. Modern reasoning models do this internally; in standard models it is activated with phrases like 'think step by step and show your reasoning'.
Temperature
An API parameter that controls generation randomness: low values (0-0.3) produce deterministic, consistent outputs (ideal for extraction and code); high values (0.7-1) produce more creative and varied ones (ideal for brainstorming and copy). It is the dial between precision and creativity.
Development
API
An interface that lets one program communicate with another: you send a structured request and receive a response. LLMs are consumed via API (you send messages and receive text), and business integrations (Stripe, GitHub, Notion) expose APIs that MCPs turn into model tools.
n8n
Open-source workflow automation tool, an alternative to Zapier/Make with a self-hosted option. It connects hundreds of services with visual logic and AI nodes: it is the usual foundation for agent automations (webhook in → LLM processes → action out).
Zapier
No-code automation platform connecting more than 7,000 applications with simple rules (when X happens, do Y). Its AI feature lets you insert LLM steps into flows. More expensive than n8n but maintenance-free: ideal for teams without a technical profile.
Webhook
A mechanism by which one system notifies another in real time: when an event occurs (sale, form, issue), it sends an HTTP request to the URL you registered. It is the usual entry point for AI automations: something happens → webhook fires → agent processes.
Web scraping
Automated extraction of data from web pages: prices, content, catalogs. Tools like Firecrawl or Crawlee structure it for AI (converting HTML to clean markdown). With a connected scraping MCP, the model can research competitors and trends on its own.
Self-hosted
Software you run on your own server instead of the provider's cloud service. Advantages: full data control, predictable costs, customization. n8n, open models like Llama, and many AI ecosystem tools have a self-hosted version for privacy requirements.
Markdown
Plain-text format with simple notation for structure (# headings, **bold**, lists). It is the de facto standard of the AI ecosystem: LLMs generate it and understand it better than HTML, and files like CLAUDE.md, SKILL.md, and GitHub READMEs use this notation.
Business
Monthly Recurring Revenue (MRR)
Monthly recurring revenue: the key metric for subscription businesses. Turning a one-time sale into a subscription multiplies customer value over time (LTV) and gives cash-flow predictability. That's why digital product models add monthly plans alongside one-time payment.
Lead magnet
A free, valuable resource (pack, template, guide, tool) delivered in exchange for an email to build a subscriber list. A subscriber who downloads something useful trusts faster: the resulting newsletter is the sales channel with the best ROI in digital products.
Affiliate program
A commission system for recommending products: each affiliate gets a tracked link and earns a percentage of the sales it generates (in AI, 20% to 50% recurring is typical). It allows growth with creators' traffic without fixed marketing costs.
Programmatic SEO
Creation of optimized pages at scale combining templates with structured data: 100 niches × 4 resource types = 400+ pages capturing low-competition long-tail searches. HubSpot and Zapier built their early traffic with this strategy; with AI, each page's content is personalized by data.
Long-tail keyword
A specific search with low individual volume but high intent: 'best Claude skills for marketing' versus 'skills'. Competing for long-tail is viable for new sites because head searches are taken; programmatic SEO attacks them systematically.
Customer Lifetime Value (LTV)
The total value a customer brings throughout their entire relationship with the business: initial purchases + subscriptions + upsells. It grows with included updates, community, and new products for the same audience. The Custom Plan and recurring subscription are direct LTV levers.
Digital product
An asset created once and sold infinitely with no marginal cost: packs, templates, courses, software. The 'build once, sell many times' model of info products — combined with curation and continuous updates, it becomes an asset that appreciates over time.