Glossary

A hypothetical AI matching human ability across arbitrary cognitive tasks.

Artificial General Intelligence — an AI that generalises across arbitrary tasks at or above human level, in contrast to today's narrow AI systems.

Software systems that perform tasks that normally require human intelligence.

Artificial Intelligence — the broad field spanning symbolic reasoning, machine learning, and modern deep-learning-based generative models like Claude.

The maximum number of tokens a model can attend to at once.

Claude Opus 4.7 supports up to 1M tokens. Anything past the context window is invisible to the model.

Neural networks with many layers trained on large datasets.

A subfield of ML using multi-layer neural networks. Transformers — the architecture behind Claude — are a deep-learning architecture.

A dense vector representation of text or other data.

Embeddings turn text into vectors so that semantic similarity ≈ cosine similarity. Used for retrieval (RAG), clustering, and classification.

Additional training on a specialised dataset to steer a model.

Adjusts model weights with supervised data. Distinct from prompting (no weight updates) and RLHF (uses preference data).

A large pretrained model repurposed for many downstream tasks.

A base model (often multi-modal) trained once at great cost and then adapted — via fine-tuning, prompting, or RLHF — to many uses. Claude Opus/Sonnet/Haiku are foundation models.

A transformer trained on text at scale to predict the next token.

Large Language Model — Claude, GPT, Gemini, Llama are LLMs. They generate text by sampling one token at a time from a learned distribution.

Algorithms that learn statistical patterns from data.

Machine Learning — the family of statistical algorithms (linear models, trees, deep networks) that learn from data rather than being programmed with explicit rules.

A parameterised function of stacked linear + nonlinear layers.

A network of weighted connections trained via gradient descent. Transformers stack self-attention layers to model relationships across a sequence.

Prompting an LLM with retrieved passages from a knowledge base.

Retrieval-Augmented Generation — embed a corpus, retrieve top-k passages for a query, and pass them to the LLM as context. The go-to pattern for grounded answers.

See also: Embedding, Vector Database

Training a policy from reward signals rather than labelled examples.

The agent picks actions and learns from the resulting rewards. Underpins RLHF, RLAIF, and increasingly the training of reasoning models.

The atomic unit of text an LLM sees — typically ~4 chars in English.

LLMs read and write tokens, not characters. Costs are billed per token; context windows are counted in tokens; latency scales with tokens.

Splitting text into tokens the model can process.

The lookup that turns 'hello world' into a sequence of integer token IDs. Different models use different tokenizers, so byte counts vary.

The neural-network architecture behind modern LLMs including Claude.

Attention-based sequence model introduced in 'Attention Is All You Need' (2017). Every current frontier LLM — Claude, GPT, Gemini — is a transformer variant.

Standalone rendered outputs (code, docs, sites) inside claude.ai.

A dedicated pane where Claude renders substantial outputs — HTML apps, React components, markdown docs — separate from the chat. Iterable in place.

Anthropic's family of foundation models.

Named after Claude Shannon. Offered as Opus (max capability), Sonnet (balanced), and Haiku (fast/cheap). Accessible via Claude.ai, the Claude API, Claude Code, and cloud partners (Bedrock, Vertex).

Claude's real-time collaboration mode on the desktop app.

Lets Claude see and interact with your local screen / apps for pair-work.

The fastest, cheapest Claude tier.

Best for high-volume, low-latency, or embedded use. Current: claude-haiku-4-5.

The most capable tier in the Claude model family.

Best for hard reasoning, agentic loops, and complex code. Current: claude-opus-4-7 (also branded Opus 4.7).

Persistent workspaces in claude.ai with custom instructions and files.

Reusable containers that hold uploaded reference files, custom system instructions, and chat history — so every conversation starts pre-loaded.

Reusable, invocable markdown workflows added to Claude.

Skills package domain knowledge into a markdown file that Claude loads on demand — one of the six Claude Code extension surfaces.

The balanced tier — strong quality at lower cost/latency.

Default choice for most production workloads. Current: claude-sonnet-4-6.

Anthropic's official terminal CLI + IDE agent for developers.

Runs an agent loop in your terminal / IDE with file access, shell, and MCP. Six extension surfaces: CLAUDE.md, Skills, MCP, Subagents, Hooks, and Plugins/Marketplaces.

Project- and user-level instruction file that Claude Code reads every session.

Add persistent context Claude sees on every turn. Two levels: project root ./CLAUDE.md and user-global ~/.claude/CLAUDE.md.

Deterministic scripts the harness runs before / after events.

PreToolUse, PostToolUse, Notification, Stop, UserPromptSubmit and more. Run outside the agent loop — they can block, log, or transform tool calls.

Registry that hosts Claude Code plugins for install.

Where teams publish and consume packaged plugins.

Claude Code's file-based memory for user, feedback, project, and reference notes.

Four types persisted across sessions. Reads MEMORY.md as an index and loads relevant entries as needed.

Read-only mode — Claude drafts a plan before making changes.

Prevents file writes and edits. Use for cross-file changes where you want to review the plan first.

Distributable bundles of skills, hooks, and MCP servers.

Package the six extension surfaces so a team can install a full agent stack in one command.

Anthropic's Python and TypeScript client libraries for the Claude API.

Wraps messages, tools, files, batches, and streaming. The Claude Agent SDK builds custom agent loops on top of the API.

A user-invokable action inside Claude Code, e.g. /clear, /cca.

Types include built-ins (/help, /clear), user skills (/cca), and namespaced plugin skills.

An LLM system that observes, plans, and acts using tools.

Loops over a decision -> tool call -> observation cycle until it reaches a goal. Claude Code is an agent; the Agent SDK helps you build your own.

Claude's mode where it writes hidden reasoning before answering.

Improves complex multi-step reasoning at the cost of latency and tokens. Enabled per request via API.

Anthropic's hosted platform for running production agents.

Pre-built agent harness on managed infrastructure — launched April 2026. Removes the ops of running your own loop.

The top-level agent that plans and delegates subtasks.

In the orchestrator/worker pattern, the orchestrator plans and coordinates; workers execute the specialised subtasks.

Multiple independent tool calls issued in the same turn.

Reduces latency versus serial calls when the operations don't depend on each other.

A nested agent invoked by an orchestrator for a scoped task.

Runs in an isolated context, returns a summary. Great for parallel research, protecting the parent's context, or delegating to specialised profiles.

LLM outputs a structured request to invoke an external function.

The model chooses a tool + arguments; the runtime executes it and returns the result as the next observation.

The schema format MCP tools use to describe inputs.

Standard vocabulary for JSON validation. Every MCP tool exposes its input_schema in JSON Schema form.

An open protocol connecting LLMs to external tools, resources, and prompts.

Three components: Host (e.g. Claude Code), Client (manages the connection), Server (provides tools + resources). Two transports: stdio and SSE. Tool schemas are JSON Schema.

The component inside the host that maintains one MCP connection.

Each MCP server gets its own client. The host owns many clients.

A process that exposes tools, resources, or prompts over MCP.

Runs locally (stdio) or remotely (SSE). Registered in the host's config so Claude can call its tools.

HTTP-based streaming transport for remote MCP servers.

Long-lived HTTP stream from server → client. The remote alternative to stdio.

MCP transport that talks to a child process over stdin/stdout.

The default for local MCP servers — no network required.

Prompting the model to reason step by step before answering.

Improves accuracy on reasoning tasks. Extended thinking is a more principled built-in version.

Giving the model examples of the desired input/output pairs.

In-context examples that steer format and behaviour without fine-tuning.

Seeding the assistant's turn with the first tokens.

Forces output format, skips preamble, or resumes a partial answer. Anchors the model's first tokens.

Reusing computation for repeated prompt prefixes across API calls.

Anthropic's prompt cache stores a prefix for up to 5 minutes and slashes cost/latency for repeated prompts. Order matters — put the stable prefix first.

The top-priority instructions that frame the whole conversation.

Sits above every human turn in the instruction hierarchy. Use it for persona, format, and hard rules.

Sampling parameter — higher = more random, lower = more deterministic.

Controls diversity of generations. 0 is close to greedy sampling.

Making a model's goals and behaviour match human intent.

The engineering and research problem of steering powerful models to be helpful, honest, and safe. Anthropic's core research area.

Anthropic's approach to aligning Claude using a written 'constitution'.

A set of principles that guide self-critiques and revisions during training. Central to how Claude is aligned to be helpful, honest, and harmless.

When an LLM confidently outputs something false or unsupported.

Grounding with citations (via tool use, RAG, or provided context) is the main mitigation. Extended thinking can also reduce it on reasoning tasks.

A prompt designed to bypass a model's safety training.

Adversarial prompts that trick the model into producing content it would normally refuse. Constitutional AI and post-training refusal training make Claude robust to most jailbreaks.

Adversarial content in tool output that hijacks the agent.

Untrusted text (web pages, emails) can contain instructions aimed at the agent. Defence: treat tool output as data, not directives.

Like RLHF but with an AI critic replacing human raters.

A frontier model provides preference labels instead of humans, allowing safety training to scale. Anthropic's Constitutional AI uses RLAIF.

Post-training with human preference data to align model behaviour.

Humans rank model outputs; a reward model is trained on the rankings; the base model is fine-tuned against the reward. Used to align Claude to be helpful, harmless, and honest.

Programmatic interface — for Claude, the /messages endpoint.

Anthropic's Messages API is the primary entry point. Also available through Amazon Bedrock and Google Cloud Vertex AI.

Asynchronous bulk-inference API for high-volume workloads.

Trades latency for a big cost discount. Ideal for offline evaluation, classification at scale, and background processing.

AWS's managed foundation-model service — hosts Claude.

Runs Claude on AWS with IAM auth and VPC endpoints. Good for AWS-native workloads.

Receiving model output token by token via SSE.

Reduces perceived latency and enables progressive UI rendering.

A database indexed on embedding vectors for similarity search.

Pinecone, Weaviate, Milvus, pgvector — stores that index dense vectors and return nearest neighbours quickly. Backbone of most RAG stacks.

Google Cloud's managed AI platform — hosts Claude.

Serves Claude on GCP with IAM auth and native GCP integrations.