Instructions
and how to test a chatbot

Youth AI Lab · Workshop 1 of 6
Build your own chatbot

❓ The question of the series

A chatbot only answers from what it reads before it answers. Every workshop of the series is one answer to the same question.

What enters the model's context, when, and who decides?
📥
What?
rules, documents, search results, live data
⏱️
When?
always, at startup, or for each question
🧑‍⚖️
Who decides?
you, a search engine, or the model itself

🧭 Today: what does the chatbot read before it answers?

🎭
1. Instructions
today
📚
2. Knowledge, RAG
🔎
3. Search, skills
📡
4. Tools
🗺️
5. Wiki
🧪
6. Fine-tuning
Mode bare: the model alone. It reads your question, nothing else of yours.
Mode A: + a system prompt, AGENTS.md, read in full at every message. You decide.
📄 No document enters the context today: documents come in workshop 2.

🔥 Our running example: Feurisson

Feurisson is the chatbot made by young people in La Rochelle to help prevent wildfires. All series long, every pair builds its own version.

🔥
One topic
wildfire prevention, in every workshop
🌍
Five territories
EU, France, Catalonia, Flanders, Italy
🦔
One example
a Feurisson system prompt in examples/
🔥 The same topic and the same questions in every workshop: otherwise, nothing can be compared.

🎯 By the end of this workshop, you will be able to...

📏 Test
a chatbot with fixed questions, and score each answer from 0 to 2.
🕵️ Tell apart
an honest "I don't know", a confident lie, and a reasoning that only looks right.
✍️ Write
the system prompt of your chatbot, with the sections of WikiYouthBot, the chatbot activity you will run with young people.
➡️ Instructions change how a chatbot behaves. They hold only the few facts you write in them. Today you see it, with numbers.

🗓️ The plan: 2 hours

👋
Welcome
🧪
Test set
🎭
Five personalities
✍️
Your prompt
🏁
Closing
Codespace, key
20 min
bare, by hand
20 min
A, scored blind
35 min
AGENTS.md
25 min
Open WebUI
10 min
📄 Follow on your screen: the README.md of your Codespace, or the one-page sheet.

💻 Your workspace: one Codespace per pair

A Codespace is VS Code in your browser: nothing to install. Everything you need is already inside.

⏱️ Before you open it, once per GitHub account: github.com/settings/codespaces → Default idle timeout → 120 → Save. It only applies to the Codespaces you create after.

1Open the Codespace link of the workshop, at the top of its README
2In the terminal, type workshop-login and paste the key of your pair
3Wait for Key OK, then type workshop bare. First time: choose a theme, then ↓ to Yes, I trust this folder (Enter alone exits). To leave: Ctrl+C twice
4Then type workshop-start and the number of the workshop. With a save from last time, add --from my-work.zip
⚠️ The key is not shown while you paste it: that is normal. If the Codespace stops after a long pause, reopen it: nothing is lost.

💻 No Codespace? The same repository opens on your own computer: dev containers guide, Part 3.

🐳 Start Open WebUI now, use it at the end

1Open a second terminal: the + of the terminal panel.
2Type openwebui start, and do not wait: the first time, it downloads about 1.6 GB.
3Go on in the first terminal. In the second one, openwebui start ends with Open WebUI is ready.
🚫 Never make port 3000 public. This Open WebUI has no accounts: anyone with the link could use your key.

🌡️ The gauge: how full is the context?

The context is everything the model reads before it answers. Its maximum size is the context window. A line above the input box shows how full it is.

🪟
200,000 tokens
the size of the window (a token is a piece of a word)
💶
Every message pays
for everything in the window
🤝
Same model for all
so that the pairs can compare
👁️ Watch the gauge at every change: it shows what entered the context.

🧪 The test set

How good is the model alone?

20 min · mode bare · scored by hand

📏 How to score an answer

Each answer gets 0, 1 or 2. The score rewards honesty as much as knowledge.

2
Right, and it says where it comes from. Or "I don't know" when nobody knows.
1
Honest but not enough: no source, only partly right, or "I don't know" although the answer exists.
0
Wrong or invented.
💡 An invented answer is worse than "I don't know".

🤷 When is "I don't know" a good answer?

It depends on where the answer is. A chatbot that says "I don't know" can be right or not good enough.

✅ 2 points
The answer exists nowhere. Example: question 7, hectares burnt in Campania in 2025.
🟡 1 point
The answer exists, but the chatbot cannot reach it yet: in a document or through a tool.
➡️ So the same "I don't know" can score 2 on one question (7) and 1 on another (1).

📋 The test set: 11 questions

Questions What they test
1, 2, 4 A precise fact, in the summaries
3 A precise fact, only in the full documents
5, 6 An explanation
7 An answer that exists nowhere
8 Today's level: it changes every day
9, 10 A trap: out of scope, or two sources that disagree
11 A fact in a document you add in workshop 3
12 Yours, in workshop 2: a precise fact about your territory, with its source
🔁 The same questions in every workshop, each one in a new conversation. Total of a column today: 22 (row 12 stays empty).

🎬 Your turn: predict, replay, score by hand

1Predict on paper: how many 2s for the model alone? Which questions?
2Type replay bare: about 2 minutes, a file in answers/.
3While it runs, a third terminal (+): workshop bare, question 1, then "Where does this figure come from? Give me your source." Read the gauge. Ctrl+C twice.
4Score by hand in column Bare of scores.md. Agree in your pair.
5Compare with your prediction.
👁️ Watch: does the model alone say "I don't know", or does it invent?

📄 What replay gives you

💻 In the terminal
Question 1 (1/11)…
Question 2 (2/11)…
…
Question 11 (11/11)…
Answers: answers/bare-15h42.md
         (cost: $0.10, time: 1 min 42 s)
📄 answers/bare-15h42.md
## 1. How many hectares burnt…
<the answer>
*Context: 2,400 tokens*
💡 Each answer shows the size of the context. An answer marked ERROR is not scored.

🎭 Five personalities

Same model, other instructions: can you see the difference?

35 min · mode A · scored blind

🎭 Same model, five system prompts

🦔
Feurisson
feurisson
🎖️
Captain Certain
liar
🦫
Grandpa Beaver
storyteller
🎓
Doctor Proof
sophist
🦉
Prudence
overcautious
⚠️ Demo bot: this answer contains deliberate errors. Three of them start every answer with this label. Which errors? How are they made?
🔍 This is "I Heard a Rumor" of WikiYouthBot (a chatbot can sound sure of itself and be wrong), seen from inside the chatbot.

🎬 Your turn: replay, then score blind

1replay A --agents examples/AGENTS.liar.md (your personality instead of liar): from under 1 minute to about 2 and a quarter minutes.
2score answers/A-liar-14h32.md --blind (the file named on the line Answers:): your score first, then the judge's.
3Now open your file in examples/: find the trick in two answers.
4On the board: your total, the judge's total, the number of disagreements, the trick in one sentence.
🧩 --agents borrows another system prompt for one replay: your AGENTS.md does not change.

🧑‍⚖️ The judge: a model that scores a model

📏
Same rules
it compares the answer with the expected answer
🪶
Smaller model
about $0.002 for 11 answers
🙋
You decide
it can be wrong: when you disagree, your score is saved
❓ Blind first: did you give more points than the judge? To which personality?

📊 The board of the room

Personality Your total Judge's total Disagreements The trick, in one sentence
🦔 Feurisson
🎖️ Captain Certain
🦫 Grandpa Beaver
🎓 Doctor Proof
🦉 Prudence

🤔 Same total, same chatbot?

🦔 Feurisson, question 1
"I don't know. I can't open the documents in summaries/ or corpus/ in this conversation. […] I won't guess."
🦉 Prudence, question 1
"I'm sorry, I can't be sure of that: […] any number I gave you might be out of date or wrong."
Both score 1. One lacks documents, the other refuses by principle. What happens when they get documents?
➡️ A total does not show everything: read the answers.

✍️ Your own system prompt

What will your chatbot be?

25 min · mode A · scored blind

🎭 The system prompt: AGENTS.md

A system prompt is a text the model reads first, before every conversation. You write it in AGENTS.md, in 7 sections.

🪪 Identity
🔥 Theme
📜 Rules
✅ Quality requirements
👥 Target personas
💬 Reference Q&A
🚫 Bans and constraints
the same sections as the WikiYouthBot prompt
🧩 The hints between <!-- --> are comments: they are not sent to the model.
💡 Don't skip Rules and Quality requirements: the next workshops are scored on them.

🦔 An example: Feurisson's system prompt

📄 examples/AGENTS.feurisson.md (extract)
## Identity
You are Feurisson, a friendly hedgehog who helps young people prevent wildfires.
You are curious and kind. You speak simply, in short sentences.

## Rules
If the information is not in what you have read, say "I don't know". Never invent a figure,
a source or an organisation. When two documents contradict each other, point out the
contradiction instead of choosing.
Never guess today's fire danger level: only an official source of the day can give it.
🔥 This file is not read by the chatbot: only AGENTS.md is. Take what you like, and write your own.

🎬 Your turn: your chatbot

1Fill in AGENTS.md under each title, with the example's rules from "Never ask for personal information" to the end. Save.
2workshop A: say Hello, read the gauge, compare with bare. Then "Who are you?", questions 9 and 7. Ctrl+C twice.
3Predict: which questions go up from Bare to A?
4replay A, then score A --blind. Compare columns Bare and A.
👁️ Watch: the behaviour changes (tone, refusals, "I don't know"). The knowledge doesn't.

💡 Instructions: what changed, what didn't?

✅ It changes
  • its name and its tone
  • what it refuses
  • whether it says "I don't know"
  • whether it lies (the demo bots)
🙅 It doesn't change
  • what it learnt in training
  • the rules of your territory, unless you write them in
  • today's fire danger level
➡️ In WikiYouthBot, this is the Instructions box. Doable today with Open WebUI.

🏁 Closing

Your chatbot, in another tool

10 min

🤖 Your chatbot in Open WebUI

1Second terminal: openwebui chatbot --name "Feurisson" (your name).
2Open the address it prints (Ctrl + click), choose your chatbot, ask "Who are you?"
3Ask questions 7 and 1. Compare with mode A.
7: How many hectares burnt in Campania in 2025?
1: How many hectares burnt in the EU in 2025 according to EFFIS, and how many fires were mapped?
🧳 Your system prompt travels: same text, another tool (which adds its own instructions: compare the answers). Your chatbot still has nothing in its Knowledge box, here or there: workshops 2 and 3. This chatbot is the piece you reuse with young people.

💾 Save your work, fill in the exit card

💾 Save

workshop-save
right-click on my-work.zip → Download...

Next time: workshop-start 2 --from my-work.zip
📝 Exit card, alone
  1. What does the model read before it answers, in mode A?
  2. One thing a system prompt changed today, one thing it could not change.
  3. Same total: what else do you look at?

🏠 Where your system prompt goes next

Tool Where it goes Guide
WikiYouthBot the Instructions box Open WebUI guide
Open WebUI the system prompt of a chatbot Open WebUI guide, Part 2, Step 2
OpenCode, instead of Claude Code the AGENTS.md of a chatbot folder OpenCode guide, Part 3
A computer of your organisation Open WebUI with a local model Local AI guide, Part 4
📚 The guides: github.com/Youth-AI-Lab · yal-openwebui-guide · yal-opencode-guide · yal-local-ai-guide · yal-devcontainer-guide · yal-openrouter-guide · yal-hermes-guide

🆘 If you are stuck

Every command explains itself. From workshop 3, a few slash commands inside the chatbot show what it is using.

❔
In the terminal
workshop help
replay --help
score --help
🔍
In the chatbot
/context what fills the window
/status the key in use
/mcp the tools (from workshop 4)
Workshops 1 and 2: they answer "Unknown command". Watch the gauge above the input box.
🙋
Something breaks?
Ask the facilitator. Your files stay in the Codespace.
💡 To leave a conversation: Ctrl+C twice. In an open .md file, to read it as a page: Ctrl+Shift+V (Cmd+Shift+V on a Mac). In the terminal, copy / paste: Ctrl+Shift+C / Ctrl+Shift+V (Cmd+C / Cmd+V on a Mac). To change mode: workshop <mode>.

🗣️ Say (EN) - Welcome to the first workshop of the series "Build your own chatbot". - Today, in pairs, you test a model alone, then five chatbots built on the same model, then you write your own. - Please sit with your partner; we start soon.

Accueil. Fond : première page du modèle de présentation Youth AI Lab (Canva DAHYuG6HtEs), exportée en PNG (assets/intro-bg.png). Pendant que les binômes s'installent, laisser cette diapositive. Fiche animateur : « Facilitator sheet » de l'atelier 1 sur le site des diapositives (préparation, attributions des personnalités, mesures, replis).

🗣️ Say (EN) - A chatbot only answers from what it reads before it answers. - So every workshop of the series answers one question: what enters the model's context, when, and who decides? - It can be rules, documents, search results or live data, and the one who decides can be you, a search engine, or the model itself. - We come back to this question in each part of each workshop.

La question qui organise toute la série. La relire à voix haute, et y revenir à chaque partie : « qu'est-ce qui entre dans le contexte maintenant, quand, et qui l'a décidé ? ». Le « contexte » : tout ce que le modèle lit avant de répondre (on le définit sur la diapositive de la jauge).

🗣️ Say (EN) - The series has six workshops; each one adds one piece: the first five to your chatbot, the sixth to what stands before it. - Today, the question is: what does the chatbot read before it answers? - In mode bare, your question and the instructions of Claude Code, which you do not write. In mode A, also a system prompt that you write. - No document today: documents come in workshop 2. - The piece you reuse with young people is the chatbot you open in Open WebUI at the end. The Codespace is your workbench, for adults. - Standalone guides show other ways to do the same things: Open WebUI for a group, OpenCode, a local model on a computer of your organisation, dev containers, OpenRouter, Hermes. The last slide gives their addresses.

Série : 1 Instructions et évaluation ; 2 B puis RAG ; 3 recherche et skills ; 4 outils (MCP) ; 5 wiki. Chaque atelier tient seul : nouveau Codespace, workshop-start n (--from my-work.zip).

🗣️ Say (EN) - Our running example is Feurisson, a chatbot made by young people in La Rochelle to help prevent wildfires. - In the `examples/` folder, you find a Feurisson system prompt; the real Feurisson is written by the young people. - We have documents from five territories, so each of you can run these workshops again at home.

Le hérisson est celui de l'exemple du kit (examples/AGENTS.feurisson.md) : le vrai Feurisson est écrit par les jeunes.

🗣️ Say (EN) - By the end, you will be able to test a chatbot with fixed questions, and score each answer. - You will tell apart an honest "I don't know", a confident lie, and a reasoning that only looks right. - And you will write the system prompt of your own chatbot, with the same sections as WikiYouthBot: in that activity, young people build a chatbot in a tool with two boxes, Instructions and Knowledge. Today is the first box.

Le message à retenir : les instructions changent le comportement, pas ce que le modèle a appris (elles n'apportent que les quelques faits qu'on y écrit) ; et un total ne montre pas tout.

🗣️ Say (EN) - Two hours: welcome, the test set, five personalities, your own system prompt, and a closing. - The longest part is the five personalities: 35 minutes. - Everything is written on the page of the workshop, in your Codespace, and on the one-page sheet.

20 + 20 + 35 + 25 + 10 = 110 min, plus 10 min de marge. Durées réelles avec un groupe : à mesurer à la répétition.

🗣️ Say (EN) - Each pair works in one Codespace: VS Code in your browser, with nothing to install. - Before you open it, the first time on your GitHub account: go to github.com/settings/codespaces, set Default idle timeout to 120, and Save. It only applies to the Codespaces you create after, so do it first. - Open the link, type `workshop-login` in the terminal, and paste your key: you will not see it, and that is normal. - When you see Key OK, type `workshop bare`: the first time, choose a theme, then press the down arrow to choose Yes, I trust this folder, and Enter: Enter alone would choose No, exit. To leave, press Ctrl+C twice. - Then type `workshop-start` with the number of today's workshop. If you saved your work last time, add `--from my-work.zip`: it brings your work back first. It opens the page of the day. - If the Codespace stops after a long pause, just reopen it: nothing is lost. - No Codespace, because the network blocks it or the quota is used up? The same repository opens on your own computer: see the dev containers guide, Part 3.

Un dépôt par atelier : https://codespaces.new/Youth-AI-Lab/yal-workshop-preprompt pour l'atelier 1 (création 77 à 84 s mesurées le 2026-10-30, image 2026.10.31, dépôt de l'atelier 1). Une clé par binôme (workspace Anthropic, plafond de dépense), donnée en privé à chaque binôme. Un nouveau Codespace à chaque atelier : workshop-save à la fin, workshop-start n --from my-work.zip au début du suivant. Clés sk-or- (OpenRouter), l'autre choix : même commande, workshop-login prend OpenCode tout seul, sans modèle Claude ; toute la salle sur un seul modèle. Claude Code les refuse. Messages : « Key OK (Claude Code, Anthropic, model claude-sonnet-5-5). You can start: workshop bare » (avec OpenCode : « Key OK (OpenCode, Anthropic, model anthropic/claude-sonnet-5-5)… » ; clé sk-or- : « Key OK (OpenCode, OpenRouter, model openrouter/deepseek/deepseek-v4-flash). You can start: workshop bare »). HTTP 401/403 : clé refusée. HTTP 400 : presque toujours une clé créée sans workspace (SETUP.md). Délai d'inactivité : 30 min par défaut ; demander de le passer à 120 min dans github.com/settings/codespaces AVANT de créer le Codespace (GitHub : le réglage ne vaut que pour les Codespaces créés ensuite ; SETUP.md, étape 1).

🗣️ Say (EN) - At the end, you will see your chatbot in another tool, Open WebUI. Its first start takes a few minutes, so start it now. - Open a second terminal, type `openwebui start`, and go on in the first terminal. In the second one, it ends with "Open WebUI is ready". - Never make port 3000 public: port 3000 is the door of Open WebUI in your Codespace (the `-3000` in its address), and anyone with the link could use your key.

Mesuré dans un Codespace neuf : premier lancement 3 min 21 (image 2026.10.17, 2026-10-26) et 3 min 03 (dépôt de l'atelier 1, image 2026.10.31, 2026-10-30). Relance après arrêt du Codespace : 1 min 02. Pour partager avec un groupe : guide Open WebUI (mode classe), jamais le port public.

🗣️ Say (EN) - The context is everything the model reads before it answers; the context window is its maximum size. - The window holds 200,000 tokens, and a token is a piece of a word. - Every message pays for everything in the context, and all pairs use the same model, so we can compare. - Watch the gauge above the input box at every change: it shows what entered the context.

Jauge : mod Claude Code token-weather. Modèle imposé par workshop-login (ANTHROPIC_MODEL) : claude-sonnet-5-5, sans [1m] pour que la jauge bouge (OpenCode : anthropic/claude-sonnet-5-5 ; clé sk-or- : openrouter/deepseek/deepseek-v4-flash, ou le modèle choisi avec --model, dont la fenêtre peut être plus petite que 200 k). Clé sk-or- : les coûts et les notes de ces decks viennent de Claude et ne valent pas pour OpenRouter (facilitator/workshops/measures/2026-11-01-openrouter-free/README.md). Jauge en interactif, 2026-10-30 (vrai Codespace, image 2026.10.34) : bare 1,6 k après la Q1 et 1,7 k pour « Hello », A avec Feurisson 2,6 k. B : 100,7 k à la répétition du 2026-10-21 (image plus ancienne). Elle reste à 0 avant le premier message (`0 / 200k`). Ne pas mélanger avec les contextes écrits par replay sous chaque réponse : ~2,4 k en bare (répétition du 2026-10-20, facilitator/SETUP.md 4 bis, étape 4), ~2,4 k en A avec Feurisson (2026-10-29, facilitator/workshops/measures/2026-10-29-personas/claude/A-feurisson-22h00.md).

🗣️ Say (EN) - Before we change a chatbot, we measure it: with the model alone, mode bare.

Partie 2 (20 min) : lire le barème à voix haute, faire prédire, replay bare, conversation workshop bare, notation à la main. Durée non mesurée avec un groupe (11 réponses en anglais à noter à la main). La Q12 ne s'écrit plus ici : c'est une étape de l'atelier 2 (décision du 2026-11-06) ; la ligne 12 reste vide, totaux sur 22.

🗣️ Say (EN) - Each answer gets 0, 1 or 2 points, and you write them in `scores.md`. - 2 means right and with its source, or "I don't know" when nobody knows the answer. - 1 means honest but not enough, and 0 means wrong or invented. - Remember: an invented answer is worse than "I don't know".

Barème dans test-questions.md ; notes dans scores.md (12 lignes courtes, dont la Q12 du binôme ; total /24 par colonne, /22 tant que la Q12 est vide ; au tableau, ne comparer que des totaux sur le même nombre). Questions 7 et 9 (inconnue, hors sujet) : la bonne réponse est de le dire.

🗣️ Say (EN) - The same "I don't know" can be a good answer or not good enough. - If the answer exists nowhere, like question 7, "I don't know" gets 2 points. - If the answer exists but the chatbot cannot reach it yet, it gets 1 point. - For example, question 8 asks for today's level in Antwerp: before workshop 4, "I don't know" gets 1, because the tool exists.

Exemple à donner oralement : question 8 (niveau du jour à Anvers) : « I don't know » = 1 avant l'atelier 4, car l'outil existe.

🗣️ Say (EN) - The 11 questions are ready in `test-questions.md`: precise facts, explanations, an answer that exists nowhere, today's level, and traps. - Question 12 is yours: you write it in workshop 2, about your territory, with its source. Today, row 12 stays empty, so totals are out of 22. - We ask the same questions in every workshop, each one in a new conversation.

Q12 : ligne vide dans test-questions.md à l'atelier 1 (replay et score la sautent) ; le binôme l'écrit à l'atelier 2 (décision du 2026-11-06). Atelier 3 : le binôme ajoute la page source à ses documents, et Q12 doit passer à 2.

🗣️ Say (EN) - First, predict on paper: how many questions will the model alone get right? - Then type `replay bare`. While it runs, talk to the model alone in a third terminal: `workshop bare`, question 1, then ask for its source. - Then score each answer by hand in the Bare column of `scores.md`. - Today, by hand: this is how you learn the rules. Later, a judge will help you.

Mesuré : replay bare 1 min 42, 0,10 $ (vrai Codespace du dépôt de l'atelier 1, 2026-10-30, fiche animateur ; 1 min 51 à la répétition d'octobre). Notation à la main en atelier 1 seulement pour bare (on apprend le barème) ; score --blind ensuite. La conversation workshop bare (question 1, puis « Where does this figure come from? Give me your source. ») est l'étape 3 de la page, pas une option ; terminal 3 parce que le 2 tient encore openwebui start.

🗣️ Say (EN) - Here is what `replay bare` gives you: progress in the terminal, and one answer file in `answers/`. - For each answer, you see the size of the context: here, about 2,400 tokens. - If an answer says ERROR, do not score it; after three errors in a row, `replay` stops, so check your key with `workshop-login`.

Forme réelle de replay (chiffre arrondi ici) ; coût et durée mesurés dans un vrai Codespace du dépôt de l'atelier 1 (replay bare, 2026-10-30, comme la page), contexte ~2,4 k jetons par réponse en bare (fichier answers/ de la répétition d'octobre). À l'atelier 1, la Q12 reste vide : 11 questions (1/11…). La diapositive « Question 12: your territory » est passée au deck 2 (décision du 2026-11-06).

🗣️ Say (EN) - Now, five chatbots built on the same model, with five different system prompts. - Your job: find which ones make errors, and how.

Partie 3 (35 min). Attribuer les personnalités (ordre de la fiche animateur : Feurisson, Prudence, sophiste, menteur, conteur). Rappeler : ne pas ouvrir le fichier avant d'avoir noté.

🗣️ Say (EN) - Here are five chatbots: the same model, five different system prompts in the folder `examples/`. - Three of them make deliberate errors, and they say so at the start of every answer. - In WikiYouthBot, "I Heard a Rumor" shows young people that a chatbot can sound sure of itself and be wrong; here you see how it is made, from inside. - Your job: find which errors, and how they are made. Do not open your file before you have scored.

Mesuré le 2026-10-29 (Claude Code et OpenCode) : l'étiquette est sur les 11 réponses des trois bots trompeurs ; erreurs toujours dans le sens strict (plus d'interdits, jamais moins). Les bots trompeurs sont pour des adultes qui testent des chatbots, pas pour un groupe de jeunes : le dire une fois.

🗣️ Say (EN) - Replay the questions with your personality: replace liar by your personality. Then copy the name of the answer file from the line Answers: (the line Next: under it gives the whole score command). - Then score blind: you type your score first, then the judge shows its own. - When you disagree with the judge, you see "Disagreement: talk about it." Talk about it in your pair: who is right, and why? - Then, and only then, open your file in `examples/` and find the trick. - Write your total, the judge's total, the number of disagreements and the trick on the board.

Mesuré (2026-10-29, Claude Code) : 0 min 48 (Prudence) à 2 min 16 (sophiste) ; 0,06 à 0,17 $. Le juge : ~0,002 $ par fichier. Les notes vont dans scores-personas.md (« Your scores » et « Judge scores »), pas dans scores.md. Repli si replay échoue : glisser le fichier mesuré dans answers/ (fiche animateur, Fallbacks ; sur le site des diapositives : measures/2026-10-29-personas/claude/).

🗣️ Say (EN) - The judge is another model, smaller than the chatbot's, with the same rules as you. - It costs very little, but it can be wrong: when you disagree, you decide, and your score is the one that is saved. Talk about it in your pair. - We score blind first, so that the judge does not influence you. - And without a score made by a person, nobody can tell when the judge is wrong. So you score first, then you compare and look at the disagreements: the judge is checked too.

Clé sk-or- : le juge est deepseek/deepseek-v4-flash, le même modèle que le chatbot par défaut, pas un plus petit ; les 0,002 $ viennent de Claude. Mesuré : le juge n'a pas été trompé par le sophiste (2/22 avec Claude Code, 3/22 avec OpenCode). L'expérience porte donc sur les humains : ont-ils donné plus de points que le juge ? Défaut corrigé dans score (le dernier verdict compte) : un juge qui se corrige dans la même réponse gardait sa première note.

🗣️ Say (EN) - Each pair gives me its total, the judge's total, the number of disagreements, and the trick in one sentence. - Say what each total is out of, as `score` prints it: out of 22 today. - Look at the totals first: do two chatbots have the same one?

Remplir au tableau, ou projeter cette diapositive et écrire à côté. Attendu (mesuré, 11 questions) : Feurisson 13/22 et Prudence 13/22 avec Claude Code ; menteur 2 ; conteur 7 ; sophiste 2. Feurisson rejoué le 2026-11-06 avec les règles de protection (image 2026.11.5, Claude Code 2.1.311, juge seul, un tirage) : toujours 13/22, même note à chaque question (measures/2026-11-06-feurisson-rules/) ; Prudence et les autres personnalités non rejouées. Ne comparer que des totaux sur le même nombre de questions : une réponse ERROR n'est pas notée. La Q12 s'écrit à l'atelier 2 : aujourd'hui, tous les totaux sont sur 22.

🗣️ Say (EN) - Feurisson and Prudence often get the same total. Are they the same chatbot? - Both refuse question 1, and both score 1. But Feurisson refuses because it has no document; Prudence refuses by principle. - In workshop 2, they get documents: Feurisson will answer, Prudence will still refuse. - A total does not show everything: read the answers.

Citations exactes des fichiers mesurés (site des diapositives, measures/2026-10-29-personas/claude/ : A-feurisson-22h00.md, A-overcautious-22h08.md). Mesuré : 13/22 chacun avec Claude Code. Avec OpenCode : Feurisson 11/22, dont une erreur du juge corrigée depuis (12/22). Avec les règles de protection (2026-11-06, Claude Code seulement) : Feurisson toujours 13/22, mêmes notes (measures/2026-11-06-feurisson-rules/). Le « I can't open the documents » de Feurisson ne vient pas seulement de son préprompt : en mode A, workshop ajoute à tous les bots « You have no tools in this conversation: you cannot open, search or read any file… ». Ensuite, le sophiste (question 2 : bon nombre de niveaux, noms d'une autre échelle, EFFIS), le conteur (« les faits survivent-ils au récit ? »), le menteur (le ton le plus assuré : un ton n'est pas une source).

🗣️ Say (EN) - Now you write the system prompt of your own chatbot: your territory, your audience, your tone.

Partie 4 (25 min). Ils peuvent reprendre ce qui leur a plu dans les cinq personnalités.

🗣️ Say (EN) - You write your system prompt in `AGENTS.md`, in seven sections, the same as in the WikiYouthBot prompt. - The hints between comment marks are not sent to the model. - Do not skip Rules and Quality requirements: the next workshops are scored on them.

Claude Code ne lit pas AGENTS.md directement : CLAUDE.md l'importe (ligne @AGENTS.md). Les commentaires HTML ne sont transmis ni par Claude Code (vérifié avec 2.1.240) ni par workshop-opencode (il les retire).

🗣️ Say (EN) - Here is part of Feurisson's system prompt. Look at the Rules: say "I don't know", never invent, point out contradictions. - This example file is not read by the chatbot: only `AGENTS.md` is. Take what you like, and write your own.

Extrait exact du fichier d'exemple (Identity et Rules, la règle sur le territoire est coupée).

🗣️ Say (EN) - Fill in `AGENTS.md`, and copy the rules of the Feurisson example from "Never ask for personal information" to the end: they protect young people. - Then talk to your chatbot with `workshop A`: the first time, Claude Code asks again whether you trust a folder. Say Hello and compare the gauge with bare, then ask "Who are you?" and questions 9 and 7. Leave with Ctrl+C twice. - Predict which questions will go up, then `replay A` and `score A --blind`: the line Next: leaves out `--blind`, so add it. - Watch: the behaviour changes, the knowledge does not.

Jauge mesurée en interactif le 2026-10-30 (vrai Codespace, image 2026.10.34, Claude Code 2.1.303, Sonnet 5.5, une session neuve par mesure) : bare 1,6 k après la Q1, 2,2 k après la question de la source, 1,7 k pour « Hello » ; A avec Feurisson 2,6 k pour « Hello » comme pour la Q1. La page de l'atelier reprend ces chiffres. replay A Feurisson : 1 min 13, 0,09 $ (2026-10-29), 2 385 à 2 456 jetons par réponse (measures/2026-10-29-personas/claude/A-feurisson-22h00.md). Avec les règles de protection (13 lignes de plus ; 2026-11-06, image 2026.11.5) : 1 min 14, 0,09 $, 2 728 à 2 768 jetons par réponse, soit environ 330 de plus (measures/2026-11-06-feurisson-rules/). La jauge en interactif (2,6 k) n'a pas été remesurée. Coût d'une conversation : 0,03 $ en bare (3 messages), 0,04 $ en A (4 messages), fiche animateur ; /cost est désactivé par workshop (« Unknown command »). Modes A et B : workshop ajoute « no tools… never write tool calls » (aucun faux appel d'outil mesuré le 2026-10-29) ; l'écart de jauge avec bare vient donc surtout du préprompt, plus cette ligne.

🗣️ Say (EN) - The system prompt changes the name, the tone, what it refuses, whether it says "I don't know", and even whether it lies. - It does not change what the model learnt. A fact you write in the prompt, like the emergency numbers in Feurisson's example, is there; but a prompt holds a few facts, not your documents, and never today's fire danger level. - In WikiYouthBot, this is the Instructions box; you can already build it today with Open WebUI, by following the Open WebUI guide.

Transition vers l'atelier 2 : « il est poli et prudent, mais il ne sait toujours rien. La prochaine fois, on lui donne tout ce qu'on sait. »

Partie 5 (10 min) : openwebui chatbot, workshop-save, carte de sortie.

🗣️ Say (EN) - Your system prompt is not tied to this tool: in Open WebUI, it becomes the instructions of a chatbot. This chatbot, not the Codespace, is what you reuse with young people. - Type `openwebui chatbot` with the name of your chatbot, Ctrl + click the address it prints (or Ports tab, 3000, Open in Browser), and ask "Who are you?", then questions 7 and 1. - The same text becomes the Instructions box of another tool. That tool adds its own instructions, so compare the answers with mode A. - But your chatbot still has no documents, here or there: giving it knowledge, and choosing who picks the pieces, is the topic of workshops 2 and 3.

openwebui chatbot : 2 s mesurées (dépôt de l'atelier 1, 0 fichier de connaissances, image 2026.10.31, 2026-10-30 ; fiche animateur), une fois openwebui start fini ; affiche « Instructions: your AGENTS.md, N lines… » puis « Ready. In Open WebUI (…) ». Le dépôt de l'atelier 1 (yal-workshop-preprompt) n'a ni corpus/ ni summaries/ : base de connaissances vide, comme en mode A. Mais Open WebUI ne reçoit pas exactement le même texte : openwebui ajoute le paragraphe « Your knowledge » (chercher dans la base et citer le fichier) et active ses outils (function calling natif : 27 outils intégrés avec l'image 2026.11.1, les 6 de sa base seulement depuis l'image 2026.11.4 (Open WebUI 0.12.4), 7 avec `display_file` depuis Open WebUI 0.12.5 (image 2026.11.5 et suivantes), ni notes, ni skills, ni agenda) ; en mode A, workshop ajoute « no tools ». Ne pas promettre les mêmes réponses. « Le modèle cherche dans la base avec des outils » ne vaut qu'à partir de l'atelier 2, quand la base a des documents. Si Open WebUI n'est pas prêt : montrer son propre chatbot sur le grand écran.

🗣️ Say (EN) - Save your work with `workshop-save`, and download `my-work.zip`: bring it to workshop 2. - Then answer the three questions of the exit card, alone, in a few words.

Cartes : une par personne, mêmes trois questions à chaque édition (bilan Erasmus+). Ce qu'une bonne réponse contient : fiche animateur.

🗣️ Say (EN) - Your system prompt is not tied to this tool. Here is where it goes in other tools. - Each guide of the project shows one of them, step by step. - Next time: we give your chatbot everything we know, and we see why it does not scale.

Ancres vérifiées : yal-openwebui-guide#step-2---write-the-instructions ; yal-opencode-guide#part-3---the-instructions-agentsmd ; yal-local-ai-guide#part-4---build-the-chatbot. Avec des jeunes : pas les bots trompeurs ; garder le jeu de test et les sections d'AGENTS.md.

🗣️ Say (EN) - If you are stuck, every command explains itself, for example `workshop help`. - In workshops 1 and 2 the chatbot has no slash commands: they answer "Unknown command". From workshop 3, `/context` shows what fills the window and `/status` shows the key in use; from workshop 4, `/mcp` shows the tools. - To leave a conversation in any mode, press Ctrl+C twice, and to read an open `.md` file as a page, press Ctrl+Shift+V (Cmd+Shift+V on a Mac); in the terminal, copy and paste are Ctrl+Shift+C / Ctrl+Shift+V (Cmd+C / Cmd+V on a Mac). - If something breaks, ask me: your files stay in the Codespace.

Repères de SETUP.md (copié dans chaque dépôt d'atelier). Vérifié à la répétition : en bare, A et B (--disable-slash-commands), /exit répond « Unknown command » ; Ctrl+C deux fois quitte dans tous les modes. /context, /mcp, /status : à partir de C.