A breif List of AI Keywords#
| Category | Keywords |
|---|---|
| Core AI concepts and methods | LLM, context, prompt, memory, agent, RAG, search |
| Tools & frameworks for orchestration | function calling, MCP, LangChain, skill, workflow, plugin, subagent |
| AI coding agents & platforms | Claude Code, Antigravity, Codex, Cursor, Manus, OpenClaw |
note: above list is not exhaustive, just a brief common keywords in AI field.
Explanation of Keywords#
LLM (Large Language Model)#
Language models have been around in NLP for decades. The early ones—think n-grams and small neural nets—were basically “autocomplete with a very short memory”: good at local next-word statistics, but brittle, short-context, and not great at handling anything outside familiar training patterns.
Then a bunch of practical breakthroughs landed at once: Transformers worked better, web-scale data became usable, GPUs/TPUs got faster, and training recipes became stable enough to scale. That made it feasible to train models with far more parameters. With enough data and compute, those extra parameters act like extra capacity—more room to store and compose patterns about syntax, meaning, and even bits of world knowledge. Past certain scale thresholds, performance doesn’t just improve smoothly; it can jump, and new capabilities start showing up (often called emergent abilities).
To separate these from their smaller, more limited predecessors, people started calling them Large Language Models (LLMs)—language models big enough to be broadly useful across many tasks.
Prompt#
A large language model, at its core, does just one job: predict the next token. Used as raw text continuation, it can feel unimpressive—like a fancy autocomplete.
The “magic” starts when we wrap that next-token predictor in a simple interaction format, such as prompt → response. By structuring input as a question (or instruction) and generating an answer, we turn plain continuation into a conversational interface—the first genuinely useful, semi-intelligent usage pattern.
┌───────────────────────────────────────────────────────────────────────────────────────────────┐
│ USER │
└───────────────────────────────────────────────┬───────────────────────────────────────────────┘
│
│ writes
v
┌───────────────────────────────────────────────────────────────────────────────────────────────┐
│ PROMPT │
│ (one turn input text) │
│ │
│ ┌──────────────────────────────┐ ┌──────────────────────────────────────────────┐ │
│ │ CONVERSATION │ │ CONTEXT │ │
│ │ (Q/A turns; still text) │ │ (background inside prompt) │ │
│ └──────────────────────────────┘ │ ┌──────────────────────────────────────┐ │ │
│ │ │ MEMORY │ │ │
│ │ │ (history + summaries/compress) │ │ │
│ │ └──────────────────────────────────────┘ │ │
│ └──────────────────────────────────────────────┘ │
└───────────────────────────────────────────────┬───────────────────────────────────────────────┘
│
│ USER uses PROMPT to talk to ONE proxy
v
┌───────────────────────────────────────────────────────────────────────────────────────────────┐
│ AGENT │
│ (proxy wrapper program) │
│ │
│ ┌─────────────────────────────────────────────────────────────────────────────────────────┐ │
│ │ loads SKILL (stored instructions + scripts) │ |
│ └─────────────────────────────────────────────────────────────────────────────────────────┘ │
│ ┌─────────────────────────────────────────────────────────────────────────────────────────┐ │
│ │ runs WORKFLOW (fixed/low-code pipeline) │ │
│ └─────────────────────────────────────────────────────────────────────────────────────────┘ │
│ ┌─────────────────────────────────────────────────────────────────────────────────────────┐ │
│ │ delegates SUB-AGENT (context isolation for independent subtask) │ │
│ └─────────────────────────────────────────────────────────────────────────────────────────┘ │
│ │
│ ┌─────────────────────────────────────────────────────────────────────────────────────────┐ │
│ │ calls tools via MCP (Agent ↔ Tool-services protocol) │ │
│ └─────────────────────────────────────────────────────────────────────────────────────────┘ │
│ │ │
│ v │
│ ┌──────────────────────────────┐ │
│ │ TOOL SERVICES │ │
│ │ (web search / docs&DB / vDB) │ │
│ └──────────────┬───────────────┘ │
│ │ │
│ │ retrieval pattern │
│ v │
│ ┌──────────────────────────────┐ │
│ │ RAG │ │
│ │ (retrieve → inject CONTEXT) │ │
│ └──────────────────────────────┘ │
└───────────────────────────────────────────────────────────────────────────────────────────────┘
│
│ sends PROMPT (with context/memory) / gets text
v
┌───────────────────────────────────────────────────────────────────────────────────────────────┐
│ LLM │
│ (next-token predictor) │
└───────────────────────────────────────────────┬───────────────────────────────────────────────┘
│
│ optional structured agreement for tool intent/results
v
┌───────────────────────────────────────────────────────────────────────────────────────────────┐
│ FUNCTION CALLING │
│ (LLM ↔ Agent message schema, e.g. JSON) │
└───────────────────────────────────────────────────────────────────────────────────────────────┘