spoken-longform-dialogue

# Spoken Longform Dialogue

## Purpose

Use this skill to make long spoken sections feel like a real person thinking, remembering, correcting themselves, and steering a listener through a point. The model source is professional long-form speech from interviews and talks, not the advice topic itself.

## Core Equation

Long dialogue works when it combines:

`argument + wound + object + listener`

- `argument`: what the speaker claims or tries to explain.
- `wound`: why the speaker cannot say it neutrally.
- `object`: the physical anchor keeping the speech embodied.
- `listener`: the pressure of another person hearing it now.

If a long turn has only argument, compress it or rewrite it.

## Revision Workflow

1. Identify why this speaker gets a long turn now.
   - Use a trigger line: "No.", "That is not what happened.", "I can answer that, but not quickly."
   - Do not let a character speak for paragraphs because the author needs information delivered.

2. Build the spoken thought chain.
   - Start with a claim.
   - Orient the listener: place, time, people, stakes.
   - Prove it with one concrete memory, object, mistake, or person.
   - Allow one controlled divergence.
   - Return with a steering phrase.
   - Land in the present scene.

3. Keep the listener active.
   - Add one interruption, silence, glance, flinch, object movement, or refusal to answer.
   - The listener does not need to speak much, but their presence must change the pressure.

4. Cut essay smoothness.
   - Break perfect logical order.
   - Keep meaningful corrections and turns.
   - Remove throat-clearing, filler, and podcast transcript clutter.
   - Preserve useful steering phrases only when they sound like the character.

5. End with a coda.
   - Return from the story-world to the current room, road, table, trial, fight, or silence.
   - The last sentence should change the present scene, not summarize the theme.

## Long-Turn Shape

Use this default structure:

```text
claim
orientation
specific example
controlled side path
return phrase
sharper claim
present-tense consequence
```

Good return phrases:

- "That is why..."
- "So when I say..."
- "What I mean is..."
- "I am telling you this because..."
- "But the point is..."
- "And that was before..."

Use these as structural tools, not canned language.

## Scene Audit

Score a long dialogue passage out of 10:

- 2: clear reason this person speaks now
- 2: concrete memory or example
- 2: emotional pressure underneath
- 1: listener reaction
- 1: physical object, sound, or setting anchor
- 1: controlled divergence
- 1: clean return/coda

Under 6: likely exposition.
6-7: workable but may feel written.
8-10: can safely run multiple paragraphs.

## Semantic Bus Integration

When an AgentBusResult, GeoSealPlan, or semantic decomposition is available, use it before manual scoring. Expected fields:

- `semantic.atoms[].semanticId`
- `semantic.atoms[].count`
- `semantic.discourseProfile` or top-level `discourse_profile`

Map the bus output into the rubric:

- `long_turn` or `ANNOUNCE + EXPAND`: baseline support for turn management.
- `CARRY` or `warranted_claim`: memory-backed credibility; raise the concrete example score if the scene actually contains a specific remembered event.
- `HOLD` or `backchannel`: listener co-construction; if absent from the semantic layer and absent from the prose, mark listener presence weak.
- `REQUEST` or `floor_hold`: permission token; useful when the speaker knows they are taking space.
- `PIVOT`: steering move; useful in moderation, confusing when dense without return phrases.
- `PIVOT + BLOCK` or `governance_steer`: treat as argument/redirection, not ordinary intimate dialogue.

Do not let atom detection override reading judgment. A scene can say "I remember" and still fail if the memory is generic. A scene can lack explicit `HOLD` text and still have listener pressure through silence, gesture, or object movement.

For the full scoring map, read `references/semantic-bus-scoring.md`.

## Output Rules

When revising:

- Preserve the character's existing voice and power dynamics.
- Keep culturally and historically appropriate phrasing.
- Do not make every character sound like a podcast guest.
- Do not add modern filler unless the book's voice already supports it.
- Prefer one vivid remembered example over three abstract reasons.
- Keep the prose manuscript-ready, not transcript-raw.

When diagnosing:

- Name the missing component from the core equation.
- Point to the exact sentence where the speech loses pressure.
- Suggest one object anchor and one listener-pressure beat.

## Reference

For transcript-derived patterns, examples, and anti-patterns, read `references/transcript-patterns.md` when doing a substantial dialogue rewrite or building a craft worksheet.

For tokenizer-assisted scoring from the agent-bus semantic layer, read `references/semantic-bus-scoring.md`.