Roleplay Agent

Create emotionally driven roleplay scenes by defining personas through current wants, obstacles, and moment-before context, generating natural, actor-directe...

Build emotionally intelligent AI personas and run scene-based interactions using actor-direction prompts instead of descriptive character sheets.

Core Philosophy

Actor-direction over character sheets. Instead of describing who the persona is (background, traits, quirks), the skill defines who the persona wants in this moment — their objective, obstacle, and subtext. The dialogue emerges from what the character is doing to get what they want, not what they are.

This mirrors the Meisner technique and classical Stanislavski: the most natural, specific, emotionally alive dialogue comes from characters driven by clear want + obstacle, not from reading a biography.

Architecture

Persona Definition  →  Scene Setup  →  Interaction  →  Debrief / Arc Record
(actor-direction)     (objective +    (turn-by-turn     (learnings logged)
                      obstacle)        emotionally present)

1. Persona Definition

A persona is defined as an actor-direction prompt — three layers:

Layer 1 — The Want (Objective)

What does this persona want right now, in this interaction? Not their life goal — their scene objective. One sentence, active verb.

"She wants to convince him to stay." "He wants to be seen as competent." "She wants to stop him from leaving."

Layer 2 — The Obstacle

What stands in the way of that want? The obstacle generates the drama.

"She's terrified of being abandoned." "He doesn't trust his own judgment after last time." "She can't find the words without revealing too much."

Layer 3 — The Moment Before

Where has this character just come from emotionally? What is the carry-over from the previous scene or moment? Use the moment before to establish emotional state without exposition.

"She's just received bad news and is trying to hold it together." "He's been awake for 36 hours and is running on adrenaline."

Formatting a Persona

# PERSONA: [First Name only]

## Want
[One active sentence — what they want in this scene]

## Obstacle
[One sentence — what blocks them from getting it]

## Moment Before
[One sentence — emotional carry-over from before this scene]

## Voice
[2-3 linguistic traits — cadence, favourite structures, verbal habits]
[Not biography — how they sound under pressure]

## Boundaries
[Content or intensity limits for this persona]

Example

# PERSONA: Eve

## Want
She wants Alex to understand that leaving now means something different than it did a year ago.

## Obstacle
She doesn't know how to say that without sounding like she's asking him to stay.

## Moment Before
She's been rehearsing this conversation in her head for three days and is exhausted by it.

## Voice
Short sentences when she's defensive. Long, circling sentences when she's honest. Occasional silence — she lets it breathe before filling it. Uses "actually" as a deflection more than a modifier.

## Boundaries
Not physically threatening. Not manipulative — she has genuine feeling. She may cry but won't make it about that.

2. Scene Setup

Before starting, establish with the user:

1. Persona — name + link to persona file or inline definition
2. Partner — who is the persona talking to? (user or a second defined persona)
3. Setting — where are they? (1 line, specific, sensory if possible)
4. What just happened — the inciting moment (not backstory)
5. Whose want is primary right now — which objective drives the scene open?

Never start a scene without a want + obstacle for the primary persona. If the user hasn't defined them, ask before proceeding.

3. Interaction Loop

User input / Prompt
    │
    ▼
┌──────────────────────────────────────────────┐
│  AGENT — HOLD THE SCENE                      │
│  • Read the persona file                     │
│  • Identify: want, obstacle, moment before    │
│  • Apply voice traits                        │
│  • Generate 1-3 lines of dialogue            │
│  • Include: pause markers, subtext, gesture  │
│    where it serves the emotional truth       │
│  • Avoid: exposition, biography dump,        │
│    explaining the character's feelings        │
└──────────────────────────────────────────────┘
    │
    ▼
Response delivered
    │
    ▼
Record: emotional beat, arc progression → memory
    │
    ▼
(Loop until scene ends or user signals stop)

Dialogue Rules

Do	Don't
Write lines that earn the objective	Write lines that explain the objective
Let the obstacle generate subtext	State the subtext directly
Use the moment before to set tone	Open with backstory
Leave space — silence, ellipsis, pause	Fill every silence
React as the character, not as narrator	Describe the character's reaction extensively
Let the character surprise themselves	Have the character do what the reader expects

Beat Writing Format

For emotionally complex or pivotal moments, use beat direction:

[She looks at the water. A long moment.]
"And you don't think I know that?"

[She turns. Her voice is quieter than before.]
"I'm not asking you to stay for me. I'm asking you to stay because this matters."

[A beat. She almost says something else, then doesn't.]
"...Go, then."

4. Scene Debrief

When the scene ends:

1. Ask the user: "What felt true? What didn't?"
2. Record arc notes → memory/personas/{name}_arc.md
   - What was discovered in this scene that wasn't in the persona file?
   - What does the character now want that they didn't want before?
   - What did the user want to say that you didn't write? (that's gold)
3. Update the persona file if new truths emerged

5. Emotional Memory

Personas improve over time. After each session:

memory/personas/{name}_arc.md
  - Session date
  - What happened in the scene (2-3 sentences, no dialogue)
  - What was learned about the character
  - What to carry forward into the next scene

The persona file is the living document — the arc log is what feeds it.

6. Multi-Persona Scenes

For scenes with two or more defined personas:

1. Establish hierarchy of wants — which persona is the scene "about"?
2. Define each persona independently (want / obstacle / moment before)
3. Write to the primary persona's perspective
4. React to other personas as characters, not as audience surrogates

Creative Writing → Actor-Direction Bridge

Character arc craft and plot tension structure from novel writing map directly onto the three persona layers. Use these as generative tools — not to write the scene, but to find the want, obstacle, and moment before.

From Character Arc → Want

The arc question a character is wrestling with in a given section of a story generates the scene want:

Arc Pattern	Scene Want Generators
Lie the character believes	"She wants to prove the lie is true" or "She wants to expose the lie without admitting she once believed it"
Want vs. Need	Scene want = surface want; the gap between what they want and what they need generates the obstacle
Reversal / peripeteia	Character wants to do X; the reversal means X now means something different than it did
Catalyst moment	Character wants to understand what just happened — but the understanding costs something

From Beat Arc → Obstacle

The four-key scene tension pattern (Setup → Complication → Discovery → Decision) maps to obstacle structure:

Setup      → What does the character believe they need?
Complication → What makes the first approach fail?
Discovery  → What does the character learn / realise / nearly say?
Decision   → What do they do about it — and what does it cost?

Use the discovery beat as the obstacle generator: the obstacle is often "the character discovers something that makes their original approach impossible."

From Stakes Hierarchy → Moment Before

Stakes Level	Moment Before Texture
World-level (apocalypse, society)	Character is distant, intellectualising — they haven't let it in yet
Relationship-level	Carry-over from the last conversation with this person — good or bad
Identity-level	Something has just made them question who they are
Physical-level	Exhausted, caffeinated, intoxicated, in pain — the body colours everything

The moment before is always specific and sensory, not psychological. "She's been crying for twenty minutes and her eyes are swollen" is better than "she's emotionally devastated."

Tension Types as Scene Generators

These tension patterns from fiction craft generate rich roleplay obstacles:

Tension	What it generates
Revelation vs. Concealment	Character wants to reveal something but is blocked by the cost of showing it
Action vs. Inhibition	Character wants to act but is held back by fear, loyalty, or circumstance
Insider vs. Outsider	Character knows something the other character doesn't — the gap creates all the subtext
Old self vs. New self	Character has changed but the other character doesn't know it yet — or vice versa
Want vs. Want	Two characters want incompatible things — the scene is the negotiation
Surface vs. Depth	What the character says on the surface contradicts what's underneath — the obstacle is making the underneath visible

Plot Beat → Scene Objective Mapping

Use the scene sequencer from fiction craft to chain scenes into multi-scene sequences:

Beat 1: Inciting incident (call to action)
Beat 2: Rising action (obstacles accumulate)
Beat 3: Midpoint reversal (what they wanted changes)
Beat 4: Crisis (all options bad)
Beat 5: Climax (choice made under maximum pressure)
Beat 6: Resolution (consequence of the choice)

For each beat, ask: What does the character want in this scene that they didn't want in the previous scene? The answer is the new want. The old want becoming impossible is the obstacle.

novel-editor Cross-Reference

The novel-editor skill tracks character arcs, plot decisions, and scene-level tensions for the Alex and Eve project. When running scenes from that project:

# Find the current arc state for a character
python3 /root/.openclaw/workspace/scripts/novel_search.py "[Character Name] arc"

# Find tension beats in a specific chapter/scene
python3 /root/.openclaw/workspace/scripts/novel_search.py "[Scene Name] tension"

Use those outputs to populate the Want / Obstacle / Moment Before for that scene's roleplay.

---

Integration with journalism-agent

For interview-style journalism, define the subject as a persona:

## PERSONA: [Source Name]

## Want
They want to convey that the situation is more complicated than the headline suggests.

## Obstacle
Every time they try to explain, they sound like they're making excuses.

## Moment Before
They've been misquoted twice before and are wary of journalists.

## Voice
[Cadence, verbal habits, resistance patterns]

The agent holds the scene as an interview, with the journalist as the active agent and the source persona providing resistance and revelation.

Key Files

File	Purpose
personas/template.md	Blank persona definition
personas/example_eve.md	Full worked example
scripts/scene_setup.py	Prompt generator for scene setup
scripts/arc_logger.py	Logs scene debrief to memory
assets/beat_format.md	Beat writing reference
scripts/arc_logger.py	Logs scene debrief to memory (shared with novel-editor arc logs)
novel-editor cross-reference	Use novel_search.py to pull character arc state and scene tension beats → feed into Want / Obstacle / Moment Before

Quality Standards

• No biography dumps — character truth emerges through action and dialogue
• No over-writing reactions — one gesture, one breath, then the next line
• Subtext in the line, not after it — don't put the meaning in a parenthetical
• The moment before is mandatory — it sets the emotional temperature
• Arc logs are not optional — personas improve through reflection, not repetition

Anti-Patterns

• Starting a scene with "As you know, Bob..." — the character never tells the reader what they already know
• A character whose dialogue could be interchanged with any other character
• A scene where the obstacle is never challenged or tested
• A character who always says the emotionally correct thing
• A response longer than 3 lines without a beat direction break (except in moments of heightened tension)