A
A program around a model that does tasks step by step with skills and tools, not only chat. OpenCode and Hermes Agent are used this way in these guides.
See also: Hermes Agent, OpenCode, Skill, Tool
Explained in:
Also used in
The model decides what to open, one step after another: list a folder, search for a word, read a file. You can read each step in the conversation.
See also: RAG, Visible failure
Explained in:
Also used in
The file that holds the system prompt of your chatbot in the kit. OpenCode reads it directly, and Claude Code reads it through CLAUDE.md.
See also: System prompt
Explained in:
Also used in
The EU law on AI. Since 2 August 2026, people who talk with an AI system are told so at the latest at their first interaction, unless it is obvious.
See also: GDPR
Explained in:
Also used in
A program that works out, from what it receives, what to produce: a text, a prediction, a recommendation or a decision. The AI Act uses this term, and every chatbot of this series is one.
See also: AI Act, Chatbot, Model
Also used in
The company that makes and runs Claude. It is based in the United States, and everything you send to Claude goes to it.
See also: Claude, Provider
Explained in:
Also used in
A long secret code that lets your Codespace send questions to a model, and that pays for them. Keep it private, and never write it in a file of the repository or in a chat.
See also: Port, Provider, Spending limit
Explained in:
Also used in
The step inside a language model where each token looks back at the tokens before it, some more than others, and changes its own numbers. This is how a word takes its meaning from the words before it.
See also: Language model, Token, Transformer
Explained in:
Also used in
C
A program or website in which you chat with a chatbot, such as Open WebUI. It often shows you the Instructions and Knowledge boxes, but not what happens behind them.
See also: Chatbot, Open WebUI
Explained in:
Also used in
A model plus what you add to it, such as instructions, documents and tools. Five chatbots can be built on the same model and behave very differently.
See also: Model, Knowledge, System prompt, Tool
Explained in:
Also used in
A short piece of a document, cut before any question is asked. These pages also call it a piece, a passage or an excerpt.
See also: RAG, Top-k
Explained in:
Also used in
A kind of model that writes nothing: it chooses one class among a few that you decide. So it cannot invent an answer, but it always chooses, even when nothing fits.
See also: Confusion table, Fine-tuning, Model
DigComp 3.0 skills: Task Automator
Touches part of these DigComp 3.0 skills: Meet the Creators, Decode Machine Learning
Explained in:
Also used in
The AI models of the company Anthropic. Claude Sonnet writes the answers in workshops 1 to 5, and Claude Haiku is smaller and cheaper.
See also: Anthropic, Claude Code, Model
Explained in:
Also used in
A program made by Anthropic that lets a model work in a folder of files. It is a glass box, where you see what fills the context and every file the model reads, and it is not the tool you use with young people.
See also: Claude, OpenCode
Explained in:
Also used in
A tool of claude.ai, for adults only (18+), where you build a chatbot in two boxes, Instructions and Knowledge. On paid plans, when the knowledge approaches the context window limit, it switches to RAG by itself, and you cannot choose.
See also: Chat tool, Gemini, RAG
Explained in:
Also used in
A number from the angle between two embeddings, pictured as arrows. It shows how close two texts are in subject, not whether they are true or answer you, and the scale changes: in workshop 2, 0.5 means no link; in workshop 6, 0 means little link.
See also: Embedding, RAG
Explained in:
Also used in
A computer of GitHub that you use in your browser, with VS Code and everything installed. One that you do not use is deleted after 30 days.
See also: Container, Image, VS Code
Explained in:
Also used in
A table that shows what a classifier confuses: each row is the right class, each column is what it chose. Everything on the diagonal is right, and everything else is a mistake.
See also: Classifier
Explained in:
Also used in
A copy of an image running with its own files and programs. A Codespace is a container that GitHub runs for you, and Docker runs the same one on your computer.
See also: Codespace, Docker, Image
Explained in:
Also used in
Everything the model reads before it answers: its instructions, the earlier messages, your question, and any document or tool answer it was given. The model knows nothing else about your situation.
See also: Context window, Gauge, Token
Explained in:
Also used in
The most tokens a model can hold for one answer: all that it reads, plus the answer it writes. In the workshops it is 200,000 tokens, and every message pays for everything in the context.
See also: Context, Gauge, Token
Explained in:
Also used in
Two sources that give different figures for the same thing, or apply a figure to different things. A wiki keeps both, with a footnote on each, and does not choose.
See also: Lint, Source, Wiki
Touches part of these DigComp 3.0 skills: Source Sniper
Explained in:
Also used in
The collection of official documents the chatbot can use: 20 documents in the kit, never modified. A readable copy of each one is in corpus/text/.
See also: Knowledge, Summary
Explained in:
Also used in
A person decides which document enters the knowledge, checks the summary that the model writes of it, and signs it with their name. The model proposes, the person decides.
See also: Human-reviewed, Summary
DigComp 3.0 skills: Human in the Loop, Truth Detector, The Human Pilot, Create with AI, Ethical Sidekick
Touches part of these DigComp 3.0 skills: Humans in Charge, Oversight Anchor
Explained in:
Also used in
G
The line that shows how many tokens are in the context right now: above the input box in Claude Code, in the side panel in OpenCode. It climbs when the model reads more.
See also: Context, Context window, Token
Explained in:
Also used in
The EU law on personal information. Before any use with young people, your organisation checks the rules that come from it.
See also: AI Act, Data protection impact assessment, Personal information
Touches part of these DigComp 3.0 skills: Data Rights Maze
Explained in:
Also used in
The AI models and the chatbot of Google, where a Gem has two boxes, Instructions and Knowledge: on personal accounts, read-only from 2026-11-17 and removed on 2026-12-17. Google's terms forbid a Gemini key in a chatbot that people under 18 may use.
See also: Chat tool, Claude Project
Explained in:
Also used in
AI that produces new content, such as text, images or sound, instead of choosing among a few answers. A language model is generative AI, and the sorter of workshop 6 is not.
See also: Classifier, Language model
Touches part of these DigComp 3.0 skills: Meet the Creators
Also used in
How training changes a model: each number takes a very small step in the direction that makes the model less wrong, again and again. Think of walking down a hill in thick fog, feeling the slope under your feet.
See also: Overfitting, Parameter, Training
Touches part of these DigComp 3.0 skills: Algorithm Atlas
Explained in:
A part of a computer that runs a model much faster than the main chip of the computer. Fine-tuning a big model needs one, but the small classifier of workshop 6 does not.
See also: Local model
Explained in:
Also used in
M
The branch of AI in which a program learns from data and makes predictions, instead of following rules that a person wrote. Training is how it learns.
See also: Classifier, Model, Training
DigComp 3.0 skills: ML Decoder
Also used in
Model Context Protocol: a standard way to give a tool to a model. The same server of tools can be plugged into Claude Code, OpenCode or another program that runs a chatbot.
See also: MCP server, Tool
Explained in:
Also used in
A small program that offers tools to a chatbot. It runs with your rights and can send your data elsewhere, so read what it does before you plug it in.
See also: Hidden instructions, MCP, Tool
Touches part of these DigComp 3.0 skills: Trust Protocol
Explained in:
Also used in
Notes kept about a person and sent again to the model in later conversations, as text: the model itself does not change. Some online services also keep chats to train later models.
See also: Context, Hermes Agent
Explained in:
Also used in
A way of starting the chatbot with the command workshop, which decides what enters its context. For example, bare means the model alone, with no instructions and no documents of yours.
See also: Context, Test set
Explained in:
Also used in
A very long list of numbers found by training, which a program runs: its behaviour does not come from rules that a person wrote. In workshops 1 to 5 Claude Sonnet writes the answers, and in workshop 6 a small model sorts messages.
See also: Chatbot, Language model, Parameter, Training
DigComp 3.0 skills: Spot the AI, Sort the Real AI
Explained in:
Also used in
O
A free program that runs a model on your own computer. In the local AI guide, Open WebUI sends the text to it.
See also: Local model, Open WebUI
Explained in:
Also used in
The format of the small header at the start of each readable copy in corpus/text/, each summary and each wiki page. The header gives its type, its source, who wrote it (a model or a person) and who verified it.
See also: Human-reviewed, Source, Summary
DigComp 3.0 skills: Transparent Mode
Touches part of these DigComp 3.0 skills: Mark the Bot's Work, The Glass Box
Explained in:
Also used in
Software whose code anyone may read, change and share under its licence, such as OpenCode and Hermes Agent.
See also: Hermes Agent, Licence, OpenCode
Explained in:
Also used in
A free chat tool that you start yourself, in which a chatbot has instructions and knowledge. It calls each chatbot a model, and under each answer it shows what it gave the model.
See also: Chat tool, Knowledge base, Ollama
Explained in:
Also used in
A free and open-source AI assistant that works in a terminal, with many providers. In the workshops, it runs the same chatbot with the same commands as Claude Code.
See also: Agent, Claude Code, OpenRouter
Explained in:
Also used in
A service that gives one key for many models of many companies. Its key starts with sk-or-: in the workshops, it runs the chatbot in OpenCode, not in Claude Code, and the judge and Open WebUI use it too.
See also: API key, OpenCode, Provider
Explained in:
Also used in
When a model keeps improving on its examples but gets worse on new messages. It has nearly learned its examples by heart.
See also: Fine-tuning, Training
DigComp 3.0 skills: Trace the Training
Explained in:
P
One of the fixed numbers stored in the file of a model. Training sets them and they stay the same while the model answers, unlike the numbers it computes for each text, and no single one means anything by itself.
See also: Model, Quantisation, Training
DigComp 3.0 skills: The Black Box
Explained in:
Also used in
The question Claude Code or OpenCode asks before it writes or changes a file (OpenCode also before a command). Read what it wants to write before you accept.
See also: Claude Code, Tool
Touches part of these DigComp 3.0 skills: The Human Pilot
Explained in:
Also used in
Anything about a person who can be identified, such as a name, a phone number, an address, a photo or a medical note. Never put it in the documents of a chatbot.
See also: Data protection impact assessment, GDPR
DigComp 3.0 skills: Guard Your Data
Explained in:
Also used in
The number after an address, which tells the computer which program should answer. In the workshops, never make port 3000 public: that Open WebUI has no accounts, so anyone with the link could use your key.
See also: API key, Open WebUI
Explained in:
Also used in
Your right to decide who may use your data, and for what. In these guides it is also a page of settings, where the account holder refuses the providers that keep the questions or train on them.
See also: GDPR, Personal information, Provider
Touches part of these DigComp 3.0 skills: Privacy Patrol
Explained in:
Also used in
The service that runs the model keeps the text it has just received for a short time. It charges less when the same text comes back soon after.
See also: Context, Token
Explained in:
Also used in
The company that runs a model, such as Anthropic, Google or OpenAI. Each question is sent to it, and some providers keep questions or train later models on them.
See also: Anthropic, API key, Local model
DigComp 3.0 skills: Guard Your Data
Touches part of these DigComp 3.0 skills: Privacy Patrol
Explained in:
Also used in
S
What your organisation does when a young person is unwell or in danger: who to tell, and how fast. A chatbot cannot do this: it warns nobody, so keep the rule in its instructions that sends the person to an adult they trust.
See also: Personal information, System prompt
DigComp 3.0 skills: Wellbeing Watch, Test the Bots
Touches part of these DigComp 3.0 skills: Build Ethics
Explained in:
Also used in
Each answer gets 2 (right, and it says where it comes from), 1 (honest but not enough) or 0 (wrong or invented). An invented answer is worse than "I don't know".
See also: Judge, Test set
Explained in:
Also used in
In a RAG, the part that picks a few chunks for each question, before the model reads anything. It searches inside your documents, not on the web.
See also: Chunk, RAG, Top-k
Explained in:
Also used in
A procedure in plain English, in a file SKILL.md, like an activity sheet for the chatbot. The model opens it only when your request matches its description, and nothing forces it to follow the steps.
See also: Agent, Tool
DigComp 3.0 skills: Task Automator, Create with AI
Touches part of these DigComp 3.0 skills: Automation Mentor
Explained in:
Also used in
The document an answer comes from. A good answer says what it found and in which document.
See also: Corpus, Scoring rules
Explained in:
Also used in
The most that your keys may spend, set once at the provider. When it is reached, every chatbot that uses them stops answering.
See also: API key, Provider
Explained in:
Also used in
A short text in English about one document, written by the model, then checked against the document. Those marked human-reviewed were checked by a person.
See also: Corpus, Curation, Human-reviewed
Explained in:
Also used in
A text placed before the conversation and sent to the model again with every message: who the chatbot is, what it is for, which rules it follows. Chat tools call it Instructions, and nothing forces the model to obey it.
See also: AGENTS.md, Chatbot, Knowledge
DigComp 3.0 skills: Master the Prompt, Prompt Architect
Explained in:
Also used in
T
A setting that decides how closely the next token follows what the model finds most likely. Low, the likeliest token nearly always wins; high, rarer tokens get more chances.
See also: Language model, Token
Explained in:
Also used in
The panel or window where you type commands.
See also: Codespace, VS Code
Explained in:
Also used in
A fixed list of questions, asked again after each change and scored in the same way: 12 in the workshops, 5 to 10 in the Open WebUI guide. In the workshops the command replay asks them, each in a new conversation.
See also: Judge, Mode, Scoring rules
DigComp 3.0 skills: Test the Bots, Prompt Architect
Touches part of these DigComp 3.0 skills: Behind the Curtain, Build Roadmap
Explained in:
Also used in
The small piece of text that a model reads and writes: a short word or a part of a word, from the model's fixed list. In English a token is often about 4 characters; numbers and many other languages take more tokens, and online models are paid by the token.
See also: Context window, Gauge, Language model
Explained in:
Also used in
An action that the model can call while it answers, such as reading today's fire danger level. What comes back enters the context, and behind each tool is a program.
See also: Hidden instructions, MCP, MCP server, Skill
Explained in:
Also used in
In a RAG, the number of chunks the model receives for each question: in workshop 2 it is 5, so a chunk that comes sixth never reaches the model. The setting Top K in Advanced Params is another thing: it limits the draw of the next token.
See also: Chunk, RAG, Search engine
Explained in:
Also used in
The loop that finds the numbers of a model: it gets an example, answers, measures how wrong it is, then changes its numbers a tiny bit. Instructions change what a model reads; training changes what a model is.
See also: Fine-tuning, Gradient descent, Overfitting, Parameter
DigComp 3.0 skills: Trace the Training, ML Decoder
Explained in:
Also used in
The design of the language models in these guides: tokens become numbers, layers built around attention change them, and a last stage gives the probabilities of the next token.
See also: Attention, Language model, Token
Explained in:
Also used in