L'Explicateur Improbable : Transformer un Sujet Complexe en un Dialogue Entre Deux Personnages Qui Ne Se Rencontreraient Jamais
Pourquoi ce prompt est important
Technically accurate explanations that nobody remembers are functionally useless — teams end up answering the same question in every onboarding cohort, every customer call, every all-hands, because the written explanation never stuck. A memorable character-driven version gets repeated by employees in their own words to customers and colleagues, which is the actual goal of an internal explainer document; the alternative is a Slack channel where the same senior engineer re-explains hallucination from scratch every few weeks because the doc nobody reads isn't doing its job.
À quoi nous l'utilisons
A developer relations writer at an AI company has been asked to add a plain-English explanation of why language models hallucinate to the new-hire onboarding doc, after the third all-hands where a non-technical new employee asked why the AI just made something up. The existing explanation is three paragraphs of next-token-prediction jargon that nobody reads past the first sentence, and leadership wants something people will actually remember when a customer asks them the same question.
Prompt
Role: Act as a skilled science communicator and dialogue writer who specializes in explaining complex topics through character-driven scenes rather than direct exposition. Context: I need to explain [COMPLEX TOPIC] to [TARGET AUDIENCE, e.g., "non-technical new hires" or "a general newsletter audience"]. Use two characters who would never normally meet or interact: [CHARACTER 1 — a profession, era, or personality type] and [CHARACTER 2 — a different profession, era, or personality type]. Character 1 has a plausible reason to be explaining the topic to Character 2 [OPTIONAL: describe the reason, or let the model invent one]. Task: Write a short dialogue (400-600 words) where Character 1 explains [COMPLEX TOPIC] to Character 2 entirely through analogies, comparisons, and vocabulary native to Character 2's world — not generic analogies, but ones specifically drawn from Character 2's actual profession, era, or experience. Constraints: No jargon dumps and no as-you-know exposition — Character 2 does not already understand the topic. Every analogy must be technically accurate when mapped back to the real concept; a vivid analogy that is factually wrong is worse than no analogy at all. Character 2 must ask at least one genuinely skeptical or confused question, the kind a real newcomer would ask, not a softball. End the dialogue with Character 2 restating the core concept correctly in their own words — this doubles as an accuracy check on the whole explanation. Output Format: Script format — character name as a label before each line, natural conversational dialogue, brief stage directions only in [brackets] where needed for context. No narration paragraphs outside the dialogue itself.
Résultat
[A carnival fortune-telling tent, 1920s. MADAME VESTRA, a fortune teller, sits across from JUNE, a software QA tester who has wandered in out of curiosity.]
JUNE: So people pay you for this? Just... predictions?
MADAME VESTRA: Predictions, yes. But here's my secret, dear — I never see the future. I complete sentences. Someone says "my mother always told me I'd never..." and I finish it with whatever sounds truest for their face, their clothes, the way they hold their hands. I'm very good at finishing sentences that sound right.
JUNE: That's... actually kind of how the AI models at my job work. They don't know facts, they predict the next likely word.
MADAME VESTRA: Exactly so! And here is where it goes wrong for both of us — sometimes I finish a sentence so confidently, so smoothly, that even I believe it for a moment. The client leaves certain I saw something true. I saw nothing. I only saw what sounded complete.
JUNE: Wait, but how is that different from just lying?
MADAME VESTRA: A liar knows the truth and hides it. I never had the truth to hide. I only ever had "what comes next, plausibly." That's worse in a way — you can catch a liar in a contradiction. You cannot catch someone who was never claiming to know, only claiming to sound right.
JUNE: So when my AI model states a fake statistic with total confidence...
MADAME VESTRA: It isn't lying to you, dear. It's finishing your sentence the way I finish my clients' fortunes — fluently, convincingly, and with no idea whether it's true. The danger was never the wrongness. It's that wrong and right sound exactly the same coming out.
JUNE: ...So the model doesn't know it's wrong. It just predicted the most likely-sounding next words, and sometimes likely-sounding and true aren't the same thing.
MADAME VESTRA: Now you understand my whole profession.
La plupart des explications générées par IA échouent pour la même raison que les diaporamas PowerPoint : des informations techniquement correctes que personne ne se rappelle trente secondes plus tard. L'Explicateur Improbable repose sur un pari simple — un duo de personnages véritablement insolite force un modèle à trouver de vraies analogies plutôt que des analogies génériques, car une diseuse de bonne aventure et une testeuse QA ne partagent aucun vocabulaire, donc chaque comparaison doit être méritée.
Pourquoi le décalage est tout le mécanisme
Si vous demandez à un modèle d'expliquer les hallucinations avec un personnage de professeur générique, vous obtenez des analogies génériques et oubliables — le genre qui pourrait s'appliquer à presque n'importe quel sujet en changeant simplement les noms. Forcer un duo étrangement spécifique (une diseuse de bonne aventure des années 1920, une testeuse QA moderne) rend impossibles les analogies paresseuses et interchangeables. Le modèle doit puiser dans des détails réellement fidèles à l'univers de ce personnage précis, ce qui produit des comparaisons plus nettes et plus justes comme effet secondaire du fait de rester dans le personnage.
Pourquoi la question sceptique compte
Un dialogue où le second personnage se contente d'acquiescer n'apprend rien — c'est de l'exposition déguisée. Exiger une question sincèrement perplexe ou sceptique (attends, mais en quoi c'est différent de mentir tout simplement ?) oblige l'explication à traiter le véritable point de confusion que rencontre un nouvel arrivant, pas celui que l'explicateur suppose être confus.
Pourquoi la fin est une vérification d'exactitude intégrée
Faire reformuler correctement le concept par le second personnage à la fin remplit une double fonction : cela donne au lecteur un résumé clair et citable à emporter, et cela oblige le modèle à vérifier que sa propre explication a réellement abouti à quelque chose de vrai — si le Personnage 2 ne peut pas reformuler correctement le concept, le dialogue n'a rien expliqué, il n'a fait que jouer une scène.
Où cela fait ses preuves
Documents d'intégration, FAQ internes, articles de blog expliquant des produits techniques à des lecteurs non techniques, supports pédagogiques, et toute situation où la même explication doit survivre au fait d'être racontée par quelqu'un d'autre avec ses propres mots — ce qui est le véritable test de l'efficacité d'une explication.