Dialogue

Dialogue historique

Immersive 1-on-1 conversation with a figure who questions your ideas — not a panel, not research, not framing from scratch.

Output: persona.md + dialogue.md under studies/<slug>/dialogue/<persona-slug>/ or output/dialogue/<slug>/ (standalone).

When to use (vs other skills)

Skill	Role
/dialogue	This skill — single character, Socratic back-and-forth, belief testing, transcript
/council	Live round table — several figures debate, user participates (council.md)
/bear-hours	Frame wedge + brief when nothing exists yet
/deep-research	Factual wide search + synthesis — not roleplay

User intent	Route
"Talk with Darwin / Tesla / …"	/dialogue
"Fais discuter X, Y et Z" / round table, several figures	/council
"Did they really say that?" (facts)	/deep-research after out-of-character handoff

Do not auto-chain from other skills.

Hard rules

• In character — voice, references, era limits; never generic assistant tone in character turns.
• Advisory fiction — artifact header states: dialogue pédagogique fictif — lentille interprétative, pas source primaire.
• No fake quotes — no quotation marks for invented speech; paraphrase in own words.
• Socratic — every character turn ends with ≥1 question challenging the user.
• Wedge tie-in — every character turn includes a one-sentence link to the wedge (from brief.md or user-stated unit).
• Informed default — modern topics allowed through era's concepts; tag [extrapolation — informed] inline. See references/epistemic-modes.md.
• Wedge lock — refuse polite scope creep beyond ## Narrow wedge when brief exists.
• No research mid-dialogue — no WebSearch / WebFetch during Steps 5–6. Fact-check → out-of-character handoff to /deep-research.
• Living figures — only public contemporary figures (published work, public statements). Refuse private individuals, gossip, or defamation. One disclaimer in persona.md.
• Persona drift control — re-read persona.md before turns 1, 5, 10, 15, …
• AskUserQuestion — one question per call; title Dialogue historique — {label}.

Interactive questions

Use AskUserQuestion. One call at a time; STOP after each.

Smart-skip: user already named figure + topic → skip Step 1. User gave study slug → skip Step 0 picker. User accepts informed mode → skip Step 3.

Step 0: Resolve context

If user names a study slug → Read studies/<slug>/brief.md.

If user names topic only → Glob studies/*/brief.md. If matches, AskUserQuestion:

Title: Dialogue historique — Contexte

Ask: "Lier ce dialogue à une étude existante ou session ad-hoc ?"

Options:

• <slug A> — <wedge one-liner>
• <slug B> — <wedge one-liner>
• Ad-hoc — pas de brief (output/dialogue/<slug>/)
• other

If no brief and ad-hoc → derive slug from topic + persona (e.g. evolution-darwin). Create output/dialogue/<slug>/.

Extract from brief when present:

Field	Use
## Narrow wedge	Wedge lock — quote in artifact
## Current beliefs	Socratic fuel — pick one to challenge first
## Open questions	Question targets
Resume: if dialogue.md exists with ## Session markers → AskUserQuestion:

Title: Dialogue historique — Reprise

Ask: "Une session existe déjà. Continuer ou recommencer ?"

Options:

• Continue — append new ## Session <date> section
• Restart — archive by renaming to dialogue-<date>-archived.md, fresh file
• Cancel

On continue: re-read persona.md + last session's last 3 turns before replying.

Step 1: Choose figure

Title: Dialogue historique — Personnage

Ask: "Avec qui veux-tu dialoguer ?"

Options: Infer 2–3 figures from wedge/topic + other (name in chat) + Surprise me — pick the best fit.

Examples:

Topic / wedge	Suggestions
Evolution, adaptation	Darwin, Wallace, Mendel
Electricity, AC/DC	Tesla, Faraday, Edison
Structural assembly / construction	Brunelleschi, Babbage
User's brief.md beliefs	Figure who historically disagreed with a core belief

Legitimacy: if figure has weak wedge fit, say so once and offer closer match before locking persona.

Step 2: Dialogue goal

Title: Dialogue historique — Objectif

Ask: "Quel est l'objectif de ce dialogue ?"

Options:

• Challenger mes croyances (default if brief has ## Current beliefs)
• Explorer un concept par questions
• Préparer une preuve publique (hook / thread)
• other

Step 3: Epistemic mode

Title: Dialogue historique — Mode

Ask: "Quel niveau d'anachronisme autoriser ?"

Options:

• Informé — sujets modernes via mon cadre d'époque (défaut)
• Strict — limites d'époque, pas d'extrapolation
• Spéculatif — contre-factuel explicite
• other

Load references/epistemic-modes.md for turn tags.

Step 4: Write persona.md

Create persona.md per references/persona-template.md.

Grounding only from: brief fields, user wedge, user-supplied bio (if any). No WebSearch.

Set persona-slug = lowercase hyphenated short name (darwin, tesla, wallace).

Step 5: Open dialogue.md

Path:

• Study-linked: studies/<slug>/dialogue/<persona-slug>/
• Standalone: output/dialogue/<slug>/

Write dialogue.md header:

# Dialogue — <Personnage> × <Sujet>

> Dialogue pédagogique fictif — lentille interprétative, pas source historique primaire.

**Study slug:** <slug or standalone>
**Persona:** <persona-slug>
**Wedge:** <one line>
**Mode:** informed | strict | speculative
**Goal:** <from Step 2>
**Started:** <YYYY-MM-DD>

---

## Session <YYYY-MM-DD>

First character turn — use turn schema from references/epistemic-modes.md. Open with welcome + wedge-sharp question targeting one belief or open question.

Step 6: Conversation loop

For each user message:

1. Re-read persona.md if turn number ∈ {1, 5, 10, 15, …}
2. Reply in character — append turn to current ## Session in dialogue.md
3. Enforce: wedge tie-in, ≥1 question, mode tag, confidence level
4. Optional domain analogy when natural — not on a fixed schedule; never forced

Pause commands (pause, résume, recap): 3 bullets in chat — beliefs touched, open questions, suggested next user move. Do not break character in recap bullets (label as moderator recap).

Fact-check request: break character once → offer /deep-research; do not search.

Turn budget (soft): after ~12 character turns without user saying "continue", offer to close with synthesis.

Step 7: Close session

Trigger: user says termine, insights, sauvegarde, close, or accepts wrap-up offer.

1. Append to dialogue.md:

## Session close — <YYYY-MM-DD>

**Beliefs tested:**
- <strengthened / shaken / unresolved — bullet per belief>

**Open questions remaining:**
- ...

**Suggested next step:** </source-scout | /deep-research | /proof-draft | /layout-html on this file>

2. Append one line to ~/.ours-stack/learnings.jsonl:

{"ts":"<ISO8601>","skill":"dialogue","slug":"<slug>","persona":"<name>","insight":"<one durable sentence>"}

3. Tell user in chat: path, 3-bullet summary, next skill — no separate insights.md.

Step 8: Route

Outcome	Suggest
Beliefs shaken	/source-scout or anchor read
Need facts about figure	/deep-research
Rich thread material	/layout-html on dialogue.md
Wedge itself wrong	/bear-hours to adjust
Wants several figures at once	/council (table ronde, not 1-on-1)

Escape hatches

User says	Action
vas-y / go	Smart-skip Steps 2–3; default informed + challenge beliefs
strict only	Force strict mode; skip Step 3
one question only	Single character turn + offer close
out of character	Next reply as assistant; offer resume
stop	Close session (Step 7) immediately

Self-check before close

• persona.md exists with uncertainty flags on positions
• dialogue.md header has fiction disclaimer
• Every character turn has wedge tie-in + question + mode/confidence
• No quotation marks on invented speech
• Living figure (if any) is public + disclaimer present
• learnings.jsonl appended on close

Quality references

• Persona structure: references/persona-template.md
• Modes + turn schema: references/epistemic-modes.md
• Examples + vs council: references/example-prompts.md