Rules in, answers out
A calculator, a tax form. Every step was written by a person.
Learn AI From zero, for everyone
Generative AI, from the first word to a working agent.
Eight short stops. Each one is a single idea and something to play with. Fifteen minutes, or jump to the stop you need. Know the basics already? Go to what an agent is and why it needs guardrails.
01 The basics
What generative AI is, and what it is not.
Classic software follows rules someone wrote. Generative AI learned patterns from a huge pile of examples, and now makes new things that fit them.
Three kinds of software: rules give the same output for the same input, machine learning sorts and scores from examples, generative AI makes new text, images and code.
Examples in, a judgement out
A spam filter. It found its own rules in millions of examples. It sorts, it does not create.
A request in, something new out
An email, a picture, a piece of code. Claude, ChatGPT and Gemini are generative AI for text.
Three things it is not
- Not a database. It stores patterns, not pages. That is how it invents, and how it gets facts wrong.
- Not a person. It feels no doubt. Right and wrong answers sound the same.
- Not up to date. Its knowledge stops at a cutoff date unless something newer is put in front of it.
02 Under the hood
How a model answers: one small piece at a time.
A language model does one thing: look at the text so far and guess the next piece. Everything impressive is that step, repeated very fast.
Sentence:
Press Run and watch the loop.
A sentence is cut into tokens, the tokens go into the model, the model picks the next token from a few candidates with probabilities, adds it to the text, and repeats.
Training
The model read a huge amount of text and practised one exercise billions of times: guess the next piece, check, adjust.
Tokens
Pieces of text, about three quarters of a word. Limits and prices are counted in them.
Inference
Using the trained model. Text in, next token out, repeat until the answer is done.
Context window
Everything the model can see at once. Fixed size. Outside it, nothing exists.
Go deeper: the network inside a model, in 3D
Drag to orbit. Hover a node.
Layers of nodes connected by lines; a signal lights up one layer after the other, from input to output.
What follows from that
Fluent and wrong
Likely is not the same as true. An invented fact or command is a hallucination.
Same question, different answers
Picking the next token involves a little chance. Temperature sets how much.
It forgets between chats
A new conversation starts empty. Any memory is text that was saved and put back.
Input decides output
Clear request and the right files beat clever tricks. Noise in the context makes answers worse.
03 The program around the model
What a harness is: the part that gives the model hands.
A model alone only returns text. The harness is the software around it that runs tools, keeps the conversation and checks the rules.
The model wants to…
The gate asks you:
Pick an action. The model only writes a request; the gate decides.
.claude/settings.json.A request leaves the model and reaches a permission gate, which lets it through (allow), waits for you (ask) or bounces it back (deny) before it can touch your files.
Runs the loop
Sends your request to the model, carries out its step, sends the result back. Repeat.
Provides the tools
Read, edit, search, run commands, fetch pages. The model names a tool; the harness runs it.
Decides what is allowed
Before each action: go, ask the human, or refuse. The only place a "no" can be enforced.
Manages the context
Chooses what the model sees, and summarises old parts when the window fills.
04 The kinds
Five kinds of harness, by how much they can reach.
Same model, very different reach. The further right, the more of your world it can touch.
Five kinds, by reach: a chat app touches nothing on your computer; an editor agent sees the open project; a terminal agent can do anything a command in that folder could; a cloud agent works on a copy of your repository on the vendor's machines; an SDK lets you build an agent with the tools you choose.
05 One harness, up close
Claude Code, in plain words.
Anthropic's harness for the Claude models: "an agentic coding tool that reads your codebase, edits files, runs commands". You describe the task; it works on your real project.
Claude Code runs in the terminal, in editors, as a desktop app and in the browser, all on the same engine with the same instruction files and settings.
Gather, act, verify
Reads what it needs, changes things, checks the result. Repeat. You can interrupt any time.
Your project and your terminal
Every file in the folder and every command you could run there. That reach is why rules and locks exist.
cd my-project, then claude
Ask "what does this project do?". Then /init writes a first CLAUDE.md. Install steps in the overview.
You: the signup form accepts empty emails. write a failing test, then fix it. Claude: wants to edit signup.test.ts ← asks first in Manual mode You: yes Claude: tests: 1 failing → edits signup.ts → tests: all passing
Go deeper: the six pieces you can add to Claude Code
CLAUDE.md
Your project's instructions, read at the start of every session. Guidance, not enforcement.
Skills
A SKILL.md under .claude/skills/: know-how or a workflow, loaded only when relevant.
Subagents
Its own loop, its own context, reports back a summary. Keeps research out of your conversation.
Hooks
Your script, run at fixed moments, every time, whatever the model decides. This is enforcement.
MCP servers
An open standard for plugging in a database, a tracker, a browser. Each one adds reach.
Plugins
Skills, hooks, subagents and MCP servers in one install. Read what is inside first.
06 Practical
Using Claude Code well: ten habits.
Most of them follow from one fact: the context window fills up fast, and quality drops as it fills.
Fill the window. When it is full, the harness compacts.
A box fills with blocks for messages, files and command output; when it is full the oldest blocks are squashed into one summary block, and the instruction file at the bottom always stays.
Give it a check it can run
Tests, a build, a screenshot. Otherwise you are the tester.
Plan before you build
Shift+Tabfor plan mode. Skip it only for one-sentence changes.Be specific
The file, the symptom, and what "fixed" looks like.
Keep CLAUDE.md short
Under 200 lines. Cut any line that prevents no mistake.
CLAUDE.md is a request
What must never happen goes in a deny rule or a hook.
Start fresh between tasks
/clear. After two failed corrections, clear and ask again.Correct early
Escstops it.Esc Escor/rewindgoes back.Still commit to git
Checkpoints cover its file edits only, never a database.
Send research to a subagent
Hundreds of file reads stay out of your conversation.
Pick a permission mode on purpose
Deny rules hold in every mode. Bypass only in a container.
Commands worth knowing
| You type | What happens |
|---|---|
/init | Writes a starter CLAUDE.md from your project. |
/context | Shows what is using space in the context window. |
/clear | Empties the context. Use it between unrelated tasks. |
/compact | Summarises the conversation to free space. |
/rewind | Restores an earlier conversation and code state. |
/permissions | Pre-approves the tools you trust. |
/model | Switches the Claude model mid-session. |
Shift+Tab | Cycles the permission mode, including plan mode. |
claude --continue | Picks up the last conversation in this folder. |
claude -p "…" | One request, no interactive screen. For scripts. |
Sources: Best practices, CLAUDE.md, Permission modes, Extend Claude Code, read on 5 October 2026. When this page and the docs disagree, the docs are right.
07 The neighbours
Other harnesses, side by side.
Every tool reads an instruction file and has a place where approvals are set. Only the names change.
| Tool | Where it runs | Instruction file it reads | Where approvals are set |
|---|---|---|---|
| Claude Code | Terminal, editor extensions, desktop app, browser | CLAUDE.md. An AGENTS.md is read by default only when there is no CLAUDE.md. |
Permission modes, plus deny, ask and allow rules in .claude/settings.json |
| Codex | CLI, IDE extension, cloud | AGENTS.md, from your home folder and from the Git root down |
Approval policy and sandbox modes. Network access is off by default. |
| Gemini CLI | Terminal | GEMINI.md by default. One setting makes it read AGENTS.md. |
Approval modes, deny rules in its policy engine, and a sandbox that is off by default |
| Cursor | Its own code editor | AGENTS.md, in the project root and in subfolders |
Run Mode (Auto-review or Allowlist), file protections, and a sandbox for commands |
| GitHub Copilot | VS Code, and a cloud agent | AGENTS.md |
Terminal auto-approve rules in VS Code. A repository firewall for the cloud agent. |
08 Vocabulary
All the jargon, in one place.
Short, honest definitions. Search, or pick a group.
No term matches that. Try a shorter word.
Models and training
- Artificial intelligence (AI)
- The broad name for software that does tasks we used to think needed human thinking. Everything below is a part of it.
- Machine learning
- Software that finds its own rules in examples, in place of rules written by a programmer.
- Neural network
- The structure most machine learning uses: layers of simple connected units, each with adjustable numbers.
- Deep learning
- Machine learning with neural networks that have many layers. The method behind today's AI.
- Generative AI
- AI that produces new text, images, sound or code, as opposed to only sorting or scoring.
- Model
- The trained result: one large file of numbers that turns an input into an output. Claude, GPT and Gemini are model families.
- LLM, large language model
- A model trained on text to predict the next token. It writes, summarises, translates and codes.
- Foundation model
- A large general-purpose model that many different products are built on.
- Transformer
- The neural network design nearly all LLMs use. It lets the model weigh every part of the text against every other part.
- Parameters, weights
- The numbers inside a model that were adjusted during training. Large models have billions of them.
- Training
- The one-time process of adjusting the parameters on a huge amount of data. Done by the model maker, not by you.
- Training data
- The text, images or code a model learned from. Its gaps and biases show up in the answers.
- Fine-tuning
- Extra training of a finished model on a smaller, specialised set of examples.
- Knowledge cutoff
- The date training data ends. The model knows nothing newer unless it is put into the context.
- Multimodal
- Able to handle more than one kind of input or output, such as text, images and audio.
- Diffusion model
- The kind of model behind most image and video generators. It starts from noise and refines it step by step.
- Open-weight model
- A model whose parameters are published, so anyone can download and run it on their own hardware.
- Benchmark
- A fixed set of tasks used to compare models. Useful, but a high score does not promise good results on your task.
- AGI
- Artificial general intelligence: a hoped-for or feared AI that matches people at most intellectual work. There is no agreed definition or test.
How a model answers
- Token
- A piece of text, roughly three quarters of a word. Models read, write and are billed in tokens.
- Inference
- Running a trained model to get an output. Every chat message is an inference.
- Prompt
- The text you give a model: your question or instruction, plus any material you attach.
- System prompt
- Instructions the product places in front of every conversation, before your first message.
- Context window
- Everything the model can see at once, measured in tokens. It has a hard limit.
- Temperature
- A setting for how much chance goes into picking each token. Low is steadier, high is more varied.
- Hallucination
- A confident, plausible, wrong answer: an invented fact, quote, command or library.
- Reasoning, thinking
- A mode where the model first writes out working steps for itself before the final answer. Slower, often better on hard problems.
- Streaming
- Showing the answer token by token as it is produced, so you do not wait for the whole reply.
- Latency
- How long you wait for a response. Bigger models and longer contexts are usually slower.
- Embedding
- A list of numbers that stands for the meaning of a piece of text, so that similar texts get similar numbers. Used for search.
Prompting and knowledge
- Prompt engineering
- Writing requests so the model does what you mean: clear goal, needed facts, examples, the format you want.
- Context engineering
- Deciding what goes into the context window and what stays out. For agents this matters more than wording.
- Zero-shot, few-shot
- Asking with no examples, or with a few examples of the answer you want. Examples usually help.
- Chain of thought
- Asking the model to work through a problem step by step before it answers.
- RAG
- Retrieval-augmented generation: first search your documents for relevant passages, then give them to the model with the question.
- Vector database
- A store for embeddings that finds the passages closest in meaning to a question. A common part of RAG.
- Grounding
- Tying an answer to sources supplied in the context, so it can be checked and is less likely to be invented.
- Structured output
- Making the model reply in a fixed format such as JSON, so another program can read the answer.
Agents and tools
- Agent
- A model plus a harness, working toward a goal by repeatedly thinking, acting through tools and looking at the result.
- Agentic loop
- The cycle an agent runs: gather context, take an action, check the result, repeat.
- Harness
- The program around the model that runs the loop, provides the tools, enforces permissions and manages the context.
- Tool, tool call, function calling
- One ability the harness offers the model, and one use of it. The model asks in text. The harness runs it.
- MCP
- Model Context Protocol: an open standard for connecting an agent to external tools and data. Each server you add is more reach.
- Instruction file
- A markdown file the agent reads at the start:
CLAUDE.mdfor Claude Code,AGENTS.mdfor most other tools. - Skill
- A packaged set of instructions or a workflow that the agent loads only when the task calls for it.
- Subagent
- An agent started by another agent for part of the job. It has its own context and returns a summary.
- Multi-agent, orchestration
- Several agents working on one job, with one of them or a script handing out the parts and collecting the results.
- Hook
- Your own script that the harness runs at a fixed moment, such as before a tool call. It runs every time, so it can enforce a rule.
- Plugin
- An installable bundle of skills, hooks, subagents and MCP servers.
- Plan mode
- A mode where the agent reads and proposes a plan but does not edit until you approve.
- Compaction
- Summarising older parts of a long conversation to free space in the context window. Early details can get lost.
- Checkpoint
- A saved snapshot of files before the agent edits them, so the change can be rewound. Not a replacement for git.
- Memory
- Notes saved between sessions and loaded back into the context. The model itself remembers nothing.
- Headless, non-interactive mode
- Running an agent from a script or a build pipeline with no person watching the screen.
- Vibe coding
- Building software by describing it to an agent and judging the result, without reading the code.
Safety and control
- Guardrail
- Any measure that keeps an AI system inside limits: a rule, a permission check, a filter, a sandbox.
- Permission mode
- A harness setting for how much it does without asking you first.
- Deny, ask, allow rule
- A permission rule the harness enforces: never run this, pause and ask me, or go ahead.
- Sandbox
- An operating-system boundary around commands: which folders they may write to and which websites they may reach.
- Prompt injection
- Text hidden in something the agent reads that tries to give it new instructions. The model cannot reliably tell it from yours.
- Jailbreak
- A prompt written to talk a model out of its built-in safety rules.
- Data exfiltration
- Data leaving where it should stay, for example an agent tricked into sending a secret key to a website.
- Human in the loop
- A design where a person approves certain actions before they run. Ask rules are the mechanical version.
- Least privilege
- Giving a program only the access its job needs, and nothing more.
- Alignment
- The work of making a model's behaviour match what its makers and users actually intend.
- Evals
- Repeatable tests of a model or an agent on tasks you care about. The AI version of a test suite.
- Red teaming
- Deliberately attacking your own AI system to find its weaknesses before someone else does.
Cost and operations
- API
- The way one program talks to another. Products reach a model through the model maker's API.
- API key
- A secret string that identifies you to an API and is billed to you. Treat it like a password. Never paste it into a chat.
- Model provider
- The company or cloud service that runs the model for you.
- Input and output tokens
- What you send and what the model writes. Both are billed, usually at different prices.
- Prompt caching
- Reusing an unchanged beginning of a prompt across requests, which makes them cheaper and faster.
- Rate limit
- A cap on how many requests or tokens you may use in a period of time.
The safety terms are explained with examples on the agents page.
Now give your agent house rules.
About two minutes. Nothing leaves your browser. Next read: what an agent is and why it needs guardrails. Official docs under external sources.