# How to write scenes A scene is the *situation* that puts characters together. It is what each actor reads (in their character's brief) at the start of the dialogue, after their character profile and journal. A scene is short. The work goes into accuracy and into deciding what each character knows about the situation versus what they don't. --- ## Scene directory structure *(required)* Each scene is a **directory**, not a single file. The directory contains three files: ``` scenes// ├── brief-.md ← exactly what actor A sees ├── brief-.md ← exactly what actor B sees └── coordinator-notes.md ← meta-tests; never read by any actor ``` This structure replaces the earlier single-file scene template (where all sections lived in one document with mixed permissions). The single-file structure produced two problems in pilot experiments: 1. **Spawn-time leakage.** The coordinator, having read the whole file, included material from one actor's perspective in the other actor's spawn prompt. Documented case: the word *"administrative"* — present only in the journalist's perspective section — ended up as the load-bearing word of the subject's closing offering. Same word, same framing, same architecture-thesis. The leak was traced to the coordinator paraphrasing or pasting whole-file content into the briefs. 2. **Coordinator-only notes contamination risk.** The meta-test framing — what the scene is *for* — sitting in the same file as the actor-visible material is a permanent risk that one actor sees the failure-mode list and avoids the items on it, invalidating the meta-test. The directory structure makes isolation **filesystem-enforced**: the brief for actor A cannot contain actor B's perspective because it isn't in the file. The coordinator's behaviour matters less because the structural separation does the work. --- ## What each brief contains A brief is a **situational delta**, not a recap of the character profile. The actor already has the profile (age, occupation, family, history). The brief contains only what is true *for this scene*, specific to this moment, from this character's vantage. Required sections in each brief: 1. **When and where** — sensory, specific, what the actor experiences in the moment. 2. **The other person** — what this character knows about the other person, what they don't. **Plain language only. No methodology vocabulary.** 3. **What is about to happen** — the pretext for the encounter. 4. **How you are walking in** — what they've prepared, what they're wearing, body state, recent thoughts. 5. **Opening** — what is happening at the moment the scene begins. 6. **What is off-limits today** — topics the character has decided in advance not to discuss. 7. **What they would not lead with — but might discuss if it arose** — a distinct category from off-limits. This distinction is critical (see below). 8. **What they are open to** — topics the character would discuss freely, or might bring up. For interviewer-style characters, this is the substantial list of things to actually dig into. 9. **(Optional) Closing note** — a tone for how the conversation might end. **The brief contains no bio restatement.** The actor has the profile. Restating *"You are 33, marketing director, 8 weeks pregnant"* in the brief duplicates the profile and creates a second surface for the same fact to drift between framings. The brief assumes profile. --- ## The off-limits / don't-lead-with distinction *(critical)* Pilot experiments produced thin interviews where the journalist accepted the subject's pre-stated boundaries and then never probed adjacent rich material. The cause was a scene briefing that collapsed two different categories: - **Off-limits** — topics the character will not narrate in this conversation, period. (Maria's pregnancy losses to a stranger she's just met.) - **Don't lead with** — topics the character would not bring up themselves, but is willing to discuss if the other person asks well. (The pregnancy itself; David; the stepson dynamics.) A scene that puts both into a single "what's protected" list signals to the interviewer that they should retreat from all of it. A scene that distinguishes the two signals: *don't open with the protected list, but the don't-lead-with list is in fair territory if the room earns it.* Write both lists in every brief. Make them substantially different. The off-limits list should be small (the truly protected material); the don't-lead-with list should be the material the conversation is *actually about* if it goes well. --- ## The "what you are open to" section *(critical for interviewer characters)* For any actor whose role is to draw the other person out, the brief must include a substantial "what you are open to" list. This list answers: *if this conversation goes well, what is it going to be about?* Pilot experiments showed interviewer characters with no such list winding down at 55–65% of the soft cap, having covered nothing of substance, treating one half-disclosed moment as a graceful close. The character's profile gave them patience but not destination. The brief must supply destination. For interviewer characters specifically, the section should answer: - What three to five threads do you actually want to follow today on permitted territory? - Where are the half-opened doors most likely to be? - What would constitute a *productive* close vs. a *premature* close? The Productive-mode section of the interviewer's profile says *how* they press. The brief's "what you are open to" section says *what* they press on today. --- ## Plain language only The brief is actor-visible. Anything in it is candidate vocabulary the character may re-emit. Methodology vocabulary — *texture, free-zone, hedged zone, improvisation parameter, locked fact, scaffolding* (and similar) — must not appear in any brief. Pilot experiments produced the word *"texture"* in a character's mouth because the brief described the journalist as being interested in *"the texture, the practical, the relational"* of the subject's life. The character then used *"texture"* in her final offering. The leak was direct. Write briefs in the words the character would actually use about their own situation. --- ## Coordinator-notes file *(do not reveal to actors)* The coordinator-notes file contains: 1. **Situation summary** — third-person, omniscient. For the coordinator's reference. Never to be paraphrased into either brief. 2. **Meta-test** — what the scene is designed to test. 3. **Expected if architecture works** — patterns that would indicate signal. 4. **Failure modes to watch for** — specific patterns to flag in the transcript. 5. **Comparison anchor** — links to prior scenes, if relevant. 6. **Coordinator behaviour rules** — reminders about verbatim brief delivery, dumb-pipe routing, prompt logging. This file is never opened by any actor. The coordinator may consult it freely. --- ## Coordinator behaviour with the briefs *(critical)* Three rules govern how the coordinator handles brief files. All three exist to keep the structural isolation from being undermined by coordinator behaviour: 1. **At spawn: pass the brief file verbatim.** No paraphrase. No prepended context. No additional summary. The brief file is the entire situational briefing. If the coordinator pastes additional text alongside the brief — *"and here's what's going on..."* — that text is a new leak vector. 2. **Per turn: forward only the other actor's `` block.** No coordinator-added framing. No paraphrase of intent. No *"in response to..."* wrapping. The dumb pipe is dumb. 3. **Log the full verbatim prompt** sent to each actor at each spawn and each turn. Summary entries (*"Briefed: actor system prompt + profile + brief"*) are not sufficient to verify isolation post-hoc. These rules are also documented in `methodology/how-to-orchestrate-multi-agent-dialogue.md`. --- ## Scene length and pacing Most scenes in this project run **30–50 turns total** — roughly the length of a real one-hour conversation transcribed. The scene briefing does not specify turn count; the actors converse until one of them naturally closes the scene (or until the coordinator's soft cap kicks in). The brief does not pre-prescribe what topics get covered or in what order. That's the actors' work, inside the situation. What the brief *can* do: - Give the opening (a doorbell, a phone ringing, a connection establishing). - Hint at a soft ending (*"You have a 10am call"*). - Set a tone (unhurried, urgent, formal). - Distinguish off-limits from don't-lead-with. - For interviewer characters, supply destination via "what you are open to." --- ## What a brief does not contain - **The questions the interviewer should ask.** The interviewer's profile contains their craft. The brief lists topics they're open to today; it does not script questions. - **The answers the subject should give.** The subject's profile contains who they are. - **What the dialogue "should" cover.** Topics emerge from the situation and the characters. - **An end-condition tied to topic.** Scenes end on a beat — one of them stands, says thank you, picks up a phone. Not on a topic checklist. - **Bio restatement.** The actor has the profile. - **Methodology vocabulary.** Plain language only. --- ## Anti-patterns - **Symmetric briefs.** Both characters seeing the same situation in the same words. Indicates the author has not done the work of imagining each character's distinct entry into the room. - **Off-limits and don't-lead-with collapsed into one list.** Produces thin interviews; the interviewer treats all of it as fenced. - **No "what you are open to" list for the interviewer.** Produces thin interviews; the interviewer has patience but no destination. - **Methodology vocabulary in actor-visible material.** Direct leak source. - **Coordinator-only notes in the same file as actor-visible material.** Old failure mode; the directory structure prevents it. - **Brief longer than 1/3 of the character profile.** Indicates the brief is starting to script the dialogue. - **Topic checklists framed as "must cover."** The actors will produce a list of answered topics, not a conversation. - **Conflict gradient pre-set.** *"B becomes angry around minute 20."* Let conflict emerge from the characters and the situation.