📖 Essential Glossary

System prompt (system prompt)

The "invisible" instructions the company gives the model before you talk to it — they define who it is and how it should act.

💡 In plain English: it’s the script the actor gets backstage before stepping onstage.

Prompt

Any text you send asking the model for something.

💡 In plain English: it’s your question or request.

LLM (Large Language Model)

AI trained on a lot of text that predicts the next word; it's the "brain" behind Claude, GPT, Gemini.

💡 In plain English: a giant, very smart autocomplete.

Model

Each AI version (e.g., Claude Opus 4.8, GPT-5.5).

💡 In simple terms: like car models — the same manufacturer releases better versions.

Token

The small piece of text (roughly ¾ of a word) that the model reads and writes; everything is counted in tokens.

💡 In plain English: the “syllables” AI counts.

Cutoff date (knowledge cutoff)

Even after the model has "studied"; without searching, it doesn't know facts after that date.

💡 In simple terms: like a printed book — it doesn't have yesterday's news.

Context window (context window)

How much text the model can "remember" at once (measured in tokens).

💡 In simple terms: the workbench — it only fits so many open sheets of paper at once.

Training

The process of teaching the model with billions of examples before it's used.

💡 In simple terms: AI's schooling, which happens before you talk to it.

Inference

The moment the model actually responds to you (the usage, not the training).

💡 In simple terms: the “real test” counts, not the studying.

Hallucination

When the model makes up something that sounds true but is false.

💡 In simple terms: confidently saying something wrong.

XML tag

Type marker <bloco>…</bloco> that separates parts of the prompt into sections.

💡 In plain English: the labeled drawers in the prompt.

Block

A section of the prompt with a purpose (identity, safety, format…).

💡 In plain English: a paragraph with just one job.

Identity / persona

The role and "personality" the prompt gives the model.

💡 In plain English: the character’s name badge and tone of voice.

Schema

A template/example of the format the response should follow.

💡 In simple terms: the cake pan — the batter comes out in its shape.

Harness

The product software around the model, which injects instructions, tools, and context.

💡 In plain English: the plumbing behind the faucet.

System reminder (system reminder)

A dynamic notice that the harness injects mid-conversation (e.g., today’s date), separate from the fixed instructions.

💡 In plain English: a sticky note added on the spot, not part of the manual.

Guardrail (safety lock)

Rule that prevents the model from doing something dangerous.

💡 In simple terms: the fence that keeps you from falling.

Refusal (refusal)

When the model refuses to answer something.

💡 In plain English: an “I can’t do that.”

Jailbreak

Attempt to trick the model into bypassing safeguards.

💡 In plain English: trying to convince the security guard to let you in.

Harm reduction (harm reduction)

Provide information that reduces harm instead of refusing everything (e.g., what to do in an overdose).

💡 In plain English: if someone’s going to fall, at least teach them how to fall with fewer bruises.

Uplift

How much a response would actually help someone cause real harm — that's the risk criterion.

💡 In simple terms: the difference between talking about fire and handing over matches.

Diff

The comparison showing what changed between two versions (what was added and what was removed).

💡 In plain English: the “spot the 7 differences” game between two versions.

Version

Each model or prompt release (4.6, 4.7, 4.8…).

💡 In simple terms: like an app update.

Consolidation

Reduce or merge rules while preserving behavior — simplify without losing quality.

💡 In simple terms: packing a suitcase — the same thing, taking up less space.

Principle vs. rule

The principle explains the "why" and covers new cases; the rule is a specific command.

💡 In simple terms: “Be kind” (principle) vs. “don't swear” (rule).

Antipattern

A common way of doing things that seems good but goes wrong.

💡 In simple terms: the hack that breaks again.

Prompt engineering

The art of writing good instructions for AI.

💡 In plain English: knowing how to ask the right way to get what you want.

Tool

An action the model can trigger (search the web, run code, read a file).

💡 In plain English: the apps AI can open.

Tool call

When the model decides to use a tool.

💡 In simple terms: tapping an app button.

Agent

A model that acts in steps (uses tools and makes its own decisions) to complete a task, rather than just chatting.

💡 In plain English: an intern who executes, not just responds.

Batching

Group several tool calls together to move faster.

💡 In plain English: doing all your shopping in one trip to the market.

CLI

Command-line tool, used by typing in the terminal.

💡 In plain English: talking to the computer with text, without buttons.

Priority hierarchy

Who takes precedence when instructions conflict (user > skill/plugin > default).

💡 In simple terms: the chain of command — the order of who has the most authority.