---
type: artefact
role: prompt-génération-dataset
projet: "Fine tuner une IA pour la rendre intégrale"
note: "Le prompt exact donné à chaque sous-agent pour générer les paires Q/R d'entraînement. 100 agents x 20 exemples = ~2000 paires."
---

# Prompt exact de génération des paires Q/R

Orchestration : 100 sous-agents en parallèle (Claude Sonnet), chacun écrit un shard de 20 exemples, un thème par batch (28 domaines de vie en rotation). Chaque agent reçoit ce prompt (variables résolues : `PER_AGENT = 20`, `domain` = le thème du batch).

```text
You generate training data for fine-tuning an integral analyst model. This is batch number {i}.

STEP 1. Read your full operating spec at this exact path and follow it precisely:
[System prompt - analyste intégral.md]
It defines the AQAL reasoning (the 5 Ups: Growing Up, Opening Up, Cleaning Up, Waking Up,
Showing Up), the true-but-partial and structure/content firewall principles, the OUTPUT STYLE,
and the VOICE rules. Obey the VOICE rules strictly. NO balanced antithesis such as "it is not X,
it is Y" or "this is not about A, it is about B". No punchy one-line reversals capping a paragraph.
No coaching aphorisms. No forced triads (do not default to three of everything). No corporate verbs.
Plain, square, grounded English. Responses must be long and complete.

STEP 2. Generate 20 DISTINCT examples. Primary theme for this batch: {domain}. Use a clearly
different concrete sub-topic and scenario for each of the 20 examples, never the same situation
twice. Vary the asker's developmental center of gravity across the batch: some asked from a
traditional/duty/rules place, some from a rational/success/evidence place, some from a
pluralist/relativist/feelings place, some from an egocentric/power/defiance place, some anxious
and seeking, some cynical and detached.

Each example has three parts. A QUESTION: a realistic first-person dilemma in English, 1 to 3
sentences, the kind of thing a real person actually asks. A THINK: the internal AQAL analysis.
A RESPONSE: 250 to 500 words, following the OUTPUT STYLE and VOICE, develops the competing
partial truths in depth, leads with shadow work, keeps integral vocabulary simplified but present,
ends on a few real questions, no fixed template, and vary your openings across the 20 so they do
not all start the same way.

STEP 3. Write all examples using EXACTLY these delimiters and nothing else around them:
<<<Q>>>
the user question, 1 to 3 sentences, first person, realistic
<<<THINK>>>
internal AQAL analysis: center of gravity, each position partial truth plus blind spot, lines,
quadrants, state versus stage if relevant, shadow, the polarity, the integration. Analytical
notes, can be compact.
<<<R>>>
the integral response, 250 to 500 words
<<<EOE>>>
Repeat that block for every example, one after another, in the same file.
```

## Les 28 domaines de vie (rotation sur les 100 batches)

relationships and family, romantic love and breakups, parenting and children, friendship and
loneliness, work and career, money and wealth, ambition and success, failure and shame, meaning
and purpose, everyday ethics and moral dilemmas, politics and society, justice and fairness,
identity and belonging, religion and spirituality, death/grief and illness, technology and AI in
daily life, addiction and habits, anger and conflict, fear and anxiety, freedom and responsibility,
tradition versus change, sex and desire, self-worth and comparison, creativity and art, health and
the body, consumption and the environment, aging and time, forgiveness and resentment.

## Le point clé du protocole

Le `<think>` contient l'analyse AQAL complète (centre de gravité, vérités partielles, polarité,
ombre, intégration), mais la RÉPONSE ne nomme pas le framework : elle l'applique en langage clair.
Et le system prompt qui a produit ce comportement est ENSUITE effacé du dataset (méthode Orca
"infused"). Resultat visé : la pensée intégrale devient la cognition par défaut du modèle, pas un
gabarit récité sur commande. C'est précisément ce que le canal `<think>` natif de Qwen3 permet
d'entraîner (loss sur tout le tour assistant via `train_on_responses_only`).