Understand AI, understand Claude, and learn to wield it well.

The nested vocabulary

These words get used interchangeably in the press, but they nest inside each other like Russian dolls. Getting them straight is the fastest way to sound like you know what you're talking about.

For developers: what actually changed

Tokens — the atoms of the model

The context window — the model's short-term memory

Context rot — the failure mode nobody warns you about

Models, parameters & weights

The rest of the vocabulary, defined

Training pretraining → fine-tuning

Pretraining is the expensive phase: predict-the-next-token over a vast text corpus, learning language and world knowledge. Fine-tuning then shapes that raw model toward being helpful, honest, and safe on specific tasks.This is why a model has a "knowledge cutoff" — it only knows what was in its training data.

Inference

The act of running the trained model to get an output. Every time you send a prompt, that's one inference. It's what you pay for and wait on.

Temperature

A dial (roughly 0–1) on how random the model's word choices are. Low = focused and repeatable; high = varied and creative. See the demo below.Note: newer frontier models increasingly manage this internally rather than exposing a raw dial.

Hallucination

When the model produces confident, fluent text that is simply false — an invented citation, a wrong date, a fabricated API. Not lying; just pattern-completion gone off-distribution.The single biggest reason to verify AI output on anything consequential.

Prompt + system prompt

The prompt is your input. The system prompt is a higher-priority instruction set that frames the whole conversation ("You are a careful code reviewer…") and shapes tone, rules, and role.

Embeddings

Text (or images) turned into a list of numbers that captures meaning, so that similar things sit close together in mathematical space. The engine behind semantic search and retrieval.

RAG retrieval-augmented generation

Instead of relying on training memory, fetch relevant documents at query time and feed them into the context. How you get an AI to answer from your data accurately and with citations.

Tool use function calling

Letting the model call real functions — search the web, run code, query a database, send an email. The model decides when and with what arguments; your code executes and returns the result.

Agents

A model in a loop: it takes an action (via a tool), observes the result, decides the next action, and repeats until a goal is met. Claude Code is an agent for software tasks.

MCP Model Context Protocol

An open standard for connecting models to external tools and data sources through a common interface — think "USB-C for AI tools." Introduced by Anthropic, now widely adopted.

Reasoning / "thinking"

Modern models can spend extra tokens working through a problem step-by-step before answering. More thinking generally means better answers on hard tasks, at the cost of more tokens and latency.

Multimodal

A model that handles more than text — reading images, PDFs, and diagrams alongside your words. Claude's current models are multimodal on input.

Constitutional AI — Claude's distinctive training idea

The model family — matching the model to the job

Claude ships as a tiered family so you can trade capability against speed and cost. The tiers are stable ideas even as version numbers advance:

The inference pipeline, step by step

01

Tokenize

Your text is broken into tokens and mapped to numbers.

→

02

Embed

Each token becomes a vector capturing its meaning.

→

03

Attention

Every token "looks at" the others to build context.

→

04

Predict

The model scores every possible next token.

→

05

Sample & loop

Pick one, append it, and repeat from step 3.

Attention — the idea that made it all work

How a raw model becomes Claude

A

Pretraining

Learn language & world knowledge by predicting text at massive scale.

→

B

Fine-tuning

Shape it toward being a helpful assistant on real tasks.

→

C

Alignment

RLHF + Constitutional AI tune it toward helpful, honest, harmless.

→

D

Serving

Deployed behind APIs and apps for you to call.

Prompting — the one skill that transfers everywhere

A first look at the API

For developers, calling Claude is a few lines. Here's the shape of it (Python), so the pieces from earlier chapters click into place:

# pip install anthropic
from anthropic import Anthropic

client = Anthropic()  # reads your API key from the environment

message = client.messages.create(
    model="claude-sonnet-5",     # pick a tier for the job
    max_tokens=1024,              # cap the response length (in tokens)
    system="You are a concise technical writer.",  # system prompt
    messages=[
        {"role": "user", "content": "Explain context rot in two sentences."}
    ],
)

print(message.content[0].text)

The path, in five rungs

Claude Code — the expert's toolkit

This is where developers compound their leverage. The features below turn Claude Code from a chatbot in a terminal into a customizable, automatable engineering partner.

CLAUDE.md

A project file where you record how your codebase works, conventions to follow, and commands to run. Loaded automatically — it's how Claude learns your project's rules once instead of every session.

Plan mode

Have Claude research and propose a plan before it touches a single file. You approve or adjust, then it executes. The single biggest lever for quality on non-trivial tasks.

Skills

Packaged instructions and assets for a repeatable task, loaded on demand when relevant. Teach Claude your team's exact procedure for, say, writing migrations or reviewing PRs — once.

Subagents

Spawn focused helper agents for parallel or independent work — one exploring the codebase while another writes tests — each with its own clean context. Parallelism without context rot.

MCP servers

Connect Claude to external systems — GitHub, databases, your ticket tracker, internal APIs — through the Model Context Protocol. This is how Claude reaches beyond your local files.

Hooks

Shell commands the harness runs automatically at defined moments (e.g. run the linter after every edit, or block a risky command). Automation that you control, not the model.

Slash commands

Reusable prompts you invoke with /name — your own /review, /ship, or /standup. Turn a great prompt you keep retyping into a one-word command.

Memory & context hygiene

Persist durable facts across sessions, and actively manage what's in the window — compact, clear, or restart. Experts treat context as a resource to curate, not a bucket to fill.

The mindset that separates experts

Agent

A model running in an act–observe–decide loop toward a goal.

Alignment

Training a model to behave in line with human intentions and values.

Attention

The mechanism letting each token weigh the influence of every other token.

Constitutional AI

Anthropic's method of aligning a model against a written set of principles.

Context window

All the tokens a model can see in one request.

Context rot

Reliability decline as the context window fills with too much or stale content.

Embedding

A numeric vector representing the meaning of text or images.

Fine-tuning

Adapting a pretrained model toward specific behaviors or tasks.

Hallucination

Confident, fluent output that is factually wrong or fabricated.

Inference

Running a trained model to produce an output.

LLM

Large Language Model — a text-trained deep-learning model.

MCP

Model Context Protocol — open standard connecting models to tools and data.

Multimodal

Handling more than text, e.g. images and PDFs.

Parameters / weights

The learned numbers that make up a trained model.

Pretraining

The large-scale next-token phase that builds raw capability.

Prompt

The input you give the model; the system prompt frames the whole session.

RAG

Retrieval-Augmented Generation — fetch relevant docs into context at query time.

RLHF

Reinforcement Learning from Human Feedback.

Temperature

A dial on output randomness / creativity.

Token

The chunk of text a model reads and generates; the unit of context and billing.

Tool use

The model calling real functions your code executes.

Transformer

The neural-network architecture behind modern LLMs.