قاموس الذكاء الاصطناعي بالعربي الواضح: الكلمات اللي تحيرك ومثال لكل وحدةThe AI Glossary in Plain Words: The Terms That Confuse People, Each with an Example

Key terms

LLM

Large language model. A program trained on a huge amount of text that predicts the next word very well. It is the brain inside Claude, ChatGPT and Gemini.

Prompt

The request you type to the model. The clearer it is, the closer the answer gets to what you want.

Token

A piece of text. Models read and write in pieces, not words, and Arabic usually splits into more pieces than English.

Context window

How much text the model can see at once: your messages, its replies and the files you uploaded.

Agent

A model that takes a goal, breaks it into steps and uses tools until the job is done, instead of only replying.

MCP

An open standard that connects AI apps to your tools and data, like a USB-C port for AI.

The steps

1. 1

   Step 1

   Tell the app from the model

   The app is the screen you talk to: Claude, ChatGPT or Gemini. The model is what thinks and writes inside it, for example Opus 5.5 inside Claude, and GPT models inside ChatGPT.

   LLM means large language model, the kind of model inside all of these apps. Open weights means the company published the model files, so you can run it on your own machine or server.

   Why it matters: one app can offer several models, each with its own speed, price and strength.

   PromptAsk the tool about itself

   Copy

   I am new to AI. Explain in simple words the difference between the app I am talking to and the model that answers me, with an example from my work in [your field].

   Tip: When you compare two tools, compare the models inside them, not just the app names.

2. 2

   Step 2

   Write a prompt that says who you are, what you want and in what shape

   A prompt is your request. A short request gets a generic answer. A clear one gets an answer that sounds like you.

   Three things make the difference: who you are and who you write for, exactly what you want, and the shape of the answer (length, tone, bullets or a paragraph).

   Related words: instructions or system prompt, standing rules you set once that apply to every chat. Few shot means giving one or two examples of what you want before the request.

   PromptA three part prompt

   Copy

   I am [who you are, for example a cafe owner in Abu Dhabi]. I want [what you want, for example an Instagram post about Thursday's offer]. Format: [length and tone, for example three lines in a friendly local voice ending with a question]. Before you write, ask me one question if anything is missing.

   Tip: If you repeat the same instructions in every chat, put them in standing instructions or a Project.

3. 3

   Step 3

   Understand tokens, because they are the unit of everything

   Models do not read words, they read pieces called tokens. A real example from OpenAI's o200k tokenizer: the Arabic proverb «اللي ما يعرف الصقر» splits into six pieces, and the Arabic word for English takes four pieces while the word English takes one.

   Every model splits text its own way, so the numbers differ between tools, but the idea holds: Arabic usually takes more pieces.

   Why it matters: API prices are counted in tokens, and so is how much of a chat the model can see.

   PromptHave it shorten for you

   Copy

   This text is long: [paste the text]. Cut it to half the length without losing any number, name or date, and tell me what you removed.

   Tip: Long answers cost tokens and time. Ask for the exact length you need.

4. 4

   Step 4

   Know the context limit and restart smart

   The context window is how much text the model can see at once: the first message, every reply and the files. When a chat gets very long, the model pays less attention to the start and can forget details you gave early on.

   The simple fix: when answers start to drift, ask for a summary of the chat, open a new chat and paste the summary in.

   For recurring work, put your files and instructions in a Project instead of repeating them in every chat.

   PromptSummary before a new chat

   Copy

   Summarise this chat in bullets: who I am, the goal, what we agreed and the next step. Write it so I can paste it into a new chat and you can continue from it.

   Tip: One chat per topic beats one chat for everything.

5. 5

   Step 5

   Catch hallucinations before you believe them

   A hallucination is when the model says something wrong with full confidence: an invented number, a study that does not exist, a link that does not open.

   It happens because the model predicts the most likely text, it does not check it. That is why numbers, sources and dates need the most checking.

   The rule: ask for the source and open it yourself. A link that does not open, or a source that does not contain the number, means the claim is wrong until proven otherwise.

   PromptMake it separate knowledge from guesses

   Copy

   Answer [your question]. After the answer, split what you said into two parts: what you are sure of, with the source, and what is a guess or something I should check myself. If you do not know, say so.

   Tip: Grounding means tying the answer to a real source, like a file you uploaded or a web search.

6. 6

   Step 6

   Tell a chat from an agent

   A chat answers you, that is all. An agent takes a goal, breaks it into steps and uses tools until it is done. Example: arrange Thursday's meeting becomes check your calendar, write the invite and send it.

   Related words: tool use means the model uses a tool such as search or a calendar. Human in the loop means a person approves before sensitive steps.

   A good agent asks you before anything that pays, sends or deletes. Never give it more access than the job needs.

   PromptWrite a job description for your first agent

   Copy

   Help me write a job description for an agent that does [the task]. Include the goal, the tools it needs, the steps, and what it must stop and ask me before doing. Then tell me which tool suits it.

   Tip: Start with one small task you repeat every week, not a whole project.

7. 7

   Step 7

   Teach the model your way of working with a skill

   A skill is a folder with a file called SKILL.md that holds instructions, examples and sometimes scripts. The model sees only its name and description, and opens the full instructions when it needs them, so it does not fill the context for nothing.

   Example: a skill for writing captions in your style, with your rules and three past captions you liked.

   In Claude you add skills from the settings, and in Claude Code you put them in the .claude/skills folder.

   PromptPull your rules from past work

   Copy

   Here are [number] captions I wrote and liked: [paste them]. Extract the rules of my style: sentence length, tone, how I open and close, and words I repeat. Write it as a SKILL.md file with a name, a description and clear instructions.

   Tip: The skill's description is what makes the model open it at the right time, so write it clearly.

8. 8

   Step 8

   Connect AI to your tools with MCP

   MCP stands for Model Context Protocol. It is an open standard that lets apps like Claude and ChatGPT connect to your tools and data: your calendar, Notion, files, databases. Its official site compares it to a USB-C port: one way to plug anything in.

   In the apps you usually see it as connectors. Each connector gives the model new tools, like reading your calendar or adding a row to a table.

   Careful: every connector can see your data, so connect only what you need, from a source you trust.

   PromptPlan your first connector

   Copy

   I use [Claude or ChatGPT] and my daily work runs on [the tools you use]. Suggest the one connector that would save me the most time, what I could ask it once connected, and what access it takes.

   Tip: A skill teaches it how to work. MCP gives it hands to reach your tools.

9. 9

   Step 9

   More words you will hear

   API: a door for developers to talk to the model from their own software instead of the screen, billed in tokens.

   RAG: the model searches your files or sources first, then answers from them. It reduces hallucinations because the answer comes from something real.

   Fine tuning: extra training of a model on specific data so it learns a style or a task.

   Multimodal: a model that understands more than one kind of input: text, images and audio.

   Reasoning or thinking: the model takes time to work step by step before answering, useful for hard problems.

   Temperature: a number that controls randomness. Low gives steady answers, high gives more varied ones.

   PromptYour own glossary

   Copy

   I heard the word [word] in [where you heard it]. Explain it in one line with no jargon, give me an example from my work in [your field], and tell me when I need it and when I do not.

Ready prompts to copy

Ask me five short questions about these words: model, prompt, token, context window, hallucination, agent, skill, MCP. Correct each answer and explain any mistake with an example.