AI for Beginners: get good at these 8 first

In one line

Give the goal, the details, the limits, the format and what done looks like. Clear words matter more than perfect grammar.

How it works

The weak prompt leaves your AI guessing: what tone, which facts, how long, and whether it can promise a slot. The strong one answers all of that in four short sentences.

Your AI may use knowledge from training, the current chat, saved memory and available tools. Give it the facts and instructions this task needs. If the reply misses, check both your prompt and its sources. Knowing the words on this page helps you ask for a result you can check.

Try it

Rewrite your last prompt with all five parts and compare the answers. Start from this:

Goal: [what you want done]
Details: [the facts, files or audience it needs]
Limits: [length, tone, what to leave out]
Format: [list, table, short message, email]
Done: [one sentence you could check the result against]

You've got it whenYou can check the answer against your goal and limits.

In one line

An app is what you open: Claude, ChatGPT, Gemini or Claude Code. A model is the engine inside it, like Opus inside Claude, GPT Image when ChatGPT makes a picture, or Nano Banana when Gemini makes one.

How it works

Start in a chat app: you ask, it answers. Move to an agent app like Claude Code when you want the AI to work on files on your computer, from a terminal or the desktop app. The same family of models can sit behind both.

Each model has its own strengths and limits: some read images, some make them, some turn speech into text. Find them on the maker's model page or the app's model picker. On Hugging Face this page is called a model card and lists intended uses and limits.

Usage and limits are counted in tokens, small chunks of text. Anthropic's glossary puts a Claude token at about 3.5 English characters, so long documents and long chats use more of your limit.

Try it

If your app lets you choose models, read the descriptions and try the same task with two of them. Otherwise, read the maker's model descriptions. Note what each model is meant to do.

You've got it whenYou can name the app, the model and what each does best.

In one line

Your AI starts with what it learned in training. Context is the information it gets for this reply, from this chat. Memory keeps information between chats in your app or project. RAG (retrieval-augmented generation) looks up useful passages in a bigger library and adds them to the context.

How it works

Context is the information the app gives the model for this response: selected chat history, instructions, file excerpts and tool results. Memory is saved in your app or project and carries over to new chats, such as your preferences, a project's files or a CLAUDE.md in a folder. Claude and ChatGPT offer memory controls in settings when the feature is available.

RAG is a system-level lookup. It searches a larger document collection, like every policy and FAQ a business has, and hands the relevant passages to the model with your question. This lookup step does not retrain the model. Check both the retrieved passage and the answer.

Try it

Paste a public or made-up FAQ and ask for an answer with the exact supporting line. Compare that quote with the FAQ yourself. This practises answering from a supplied source; RAG adds the search step for a bigger library. Try this prompt:

Answer using only the FAQ below. After your answer, quote the line you used. If the FAQ doesn't cover it, say so.

FAQ:
[paste your FAQ here]

Question: Can I collect a birthday cake on Friday at 4pm?

You've got it whenYou can match the FAQ answer to its source line.

In one line

A .md (Markdown) file is plain text with simple formatting. Name it for its job: CLAUDE.md or AGENTS.md for rules, SKILL.md for one task, and plain knowledge files for facts.

How it works

Each file has one job, so your AI gets the right instructions at the right time. CLAUDE.md is read by Claude Code every session; Claude Code can also read AGENTS.md. A skill is a folder with a SKILL.md file plus any examples or templates, loaded when the task needs it. A knowledge file holds facts, like the bakery's FAQ, that you attach or keep in a project.

Claude Code's docs say these files are context, not enforced rules, so check the results. Some chat apps offer project instructions, custom instructions or attached files instead. Use a feature your app supports.

Try it

Write five short rules for a task you repeat and save them where your app reads them. For a project folder in Claude Code, that's CLAUDE.md. A bakery example to adapt:

# Cake replies

## Tone
Friendly and short. No emojis.

## Always ask for
- The exact collection date and time
- Cake size and the message on top

## Never
- Promise a slot. Staff confirm availability first.

## Facts
- Use cake-faq.md for notice periods and collection hours.

You've got it whenYou stop retyping the same instructions.

In one line

A workflow follows steps and branches set in advance, sometimes on a schedule. An agent chooses its next step and tools toward a goal.

How it works

Give it one specific task. Some AI apps can run a task on a schedule; developers call a timed job a cron job. A good routine starts by reading its rules file, does one job, and stops at a point where a person checks.

Then audit it. Once a week, read what it did, fix what went wrong in the rules file, and run it again. Anthropic's advice is to find the simplest solution that works and add complexity only when you need it. Start with a single prompt if it is enough; use a workflow when the job needs several defined steps.

Try it

Write the steps of one weekly task and mark where a person checks. The bakery's:

1. Each morning, read new cake enquiries.
2. Read the rules in CLAUDE.md and the facts in cake-faq.md.
3. Draft a reply to each one.
4. Staff check availability, approve and send.

You've got it whenSomeone else could follow your steps.

In one line

Grant the least access a job needs: no access, read only, draft only, or act for you. An API key is a password for software, so never share it.

How it works

Start with read only. Your AI can look in the apps you connect and tell you what's there, such as today's calendar or the newest messages, without changing anything. Move up to draft only, then to act for you, only when you've checked its work for a while, and keep a person on anything that sends, books, pays or deletes. A confident answer can still be wrong.

You set access in a few places: a connector's settings, an agent app's prompts (Claude Code can ask before it edits a file or runs a command, and you choose what to allow), and API keys. An API key lets software act as you. Keep it out of chats and screenshots, and replace it if it leaks.

Try it

If your app offers a connector with a read-only setting, use a test account or non-sensitive files. Check what it can read, then ask your AI what it can see. Keep it on read only for this exercise and find how to disconnect it. If you can't limit its access, leave it off. Write down what would need your OK.

You've got it whenNothing goes out to anyone without your check.

In one line

Connectors let your AI app reach your other apps, like your calendar or files. MCP (Model Context Protocol) is an open standard for building those links. Its docs compare it to "a USB-C port for AI applications".

How it works

A connector gives your AI app access to another service. What it can read or change depends on the connection and the permissions you grant. MCP gives app makers one shared way to build these links, the way USB-C gives devices one shared plug.

The plug and the access are separate: MCP is how the link is made, and the access level from lesson 6 decides what your AI may do through it.

Try it

Read a connector's permissions before you turn it on. If your app offers connectors, choose one you recognise and read what it may see and do before enabling it. Use a test account or non-sensitive files. If its access is too broad and cannot be narrowed, leave it off. Start it on read only where the app lets you, then ask your AI what it can see. Find how to disconnect it.

You've got it whenYou know what your AI can reach, and how to switch it off.

In one line

Git keeps save points of your project, like checkpoints in a game. GitHub stores them online, so if something breaks you can load an earlier save. It's also a free library of tools, skills and .md files that people share.

How it works

GitHub's own docs describe a commit as being like taking a photo, and say any earlier version can be recovered. Before an agent like Claude Code changes your files, make a save point. You can then restore the files you saved if a change breaks them.

GitHub is also a library. People share tools, skills (a folder with a SKILL.md file) and .md rules files there, free to download. Before you add one, read its README, check who made it and when it was last updated, and don't give it more access than the job needs.

Installing a tool often means pasting one line into Terminal from the project's README or official install page. Claude Code's install page has one. Only run commands from the official page, and read what they do first.

Try it

Use a practice project with made-up files that's already set up with Git. Ask your AI to show which files it will save, approve that list, then ask it to make a local save point and show you the history. If your app can't do this, open a public project on GitHub and look through its history instead.

You've got it whenYou can go back to an earlier save when something breaks.