How to Use AI to Write Dialogue Across a Power Imbalance

Power imbalance is one of the hardest things to write well — and one of the first things AI gets wrong. Not because the model doesn't understand hierarchy, but because it defaults to the path of least resistance: the powerful character talks, the less powerful character reacts. Lather, rinse, repeat. What you end up with is a scene where the subordinate person is basically a tennis ball, just bouncing back whatever the dominant character serves. No interiority. No agenda. No friction that feels earned.

This guide is about fixing that. Specifically, how to prompt so that both characters feel like full people — even when one of them is trapped, outranked, or outgunned.

Why Power Imbalance Breaks AI Dialogue (The Default Traps)

The first trap is what I call reflexive compliance. You write a scene where a prison warden questions an inmate, and the AI writes the inmate answering every question directly, honestly, maybe even gratefully. The warden speaks. The inmate explains himself. The warden pushes. The inmate apologizes. It reads like a customer service interaction, not a scene with any stakes.

The second trap is the opposite: cartoonish defiance. The AI overcorrects and writes the subordinate character as improbably mouthy — delivering speeches about their rights, refusing everything, practically monologuing. Real people with less power are rarely this openly combative, because they can't afford to be. Defiance in the real world is usually oblique, coded, deniable.

The third trap is narrated interiority that contradicts the spoken dialogue. The AI will write something like: "She nodded meekly, her voice barely above a whisper. 'Of course, sir,' she said, though inside she burned with rage." This is telling you she has depth while the actual words coming out of her mouth have none. The subtext is in the description, not the dialogue itself.

Real power imbalance lives in what people don't say — in deflections, half-answers, strategic agreements, and the small choices about which battles to pick. If your AI-generated dialogue isn't doing that work, the prompt is the problem.

The fix isn't just writing better prompts, though that's most of it. It's also understanding what you're actually asking the AI to do before you type anything.

Mapping the Power Geometry Before You Prompt

Before you write a single prompt, you need to know three things about your scene's power structure — and they're not as obvious as "who's the boss."

Structural power is the formal, visible stuff: job title, legal authority, financial control. A parent controls the car keys. A parole officer controls someone's freedom. A wealthy client controls the contract. This is the iceberg above water.

Relational power is messier. It's who knows what secret, who has the emotional leverage, who the other person desperately wants approval from. A daughter might have less structural power than her controlling father, but she has relational power if she's the only family he has left. That shifts everything.

Informational power is who knows something the other person doesn't. This is huge in interrogation scenes, negotiation scenes, and any scene where one character is hiding something. The character who appears to have less power might actually be holding the most important card in the room.

Map all three before you prompt. Actually write it down, even in a quick note. Something like:

• Structural power: The boss holds it (can fire her)
• Relational power: The employee holds it (she knows about the embezzlement)
• Informational power: Split — he suspects she knows, but doesn't know what she knows

When you hand this map to the AI in your prompt, you're not just describing the scene. You're giving it the architecture of the tension. The AI can then generate dialogue where the employee answers questions carefully — not out of deference, but because she's deciding in real time how much to reveal. That's a character with agency, not a doormat.

Prompting Techniques: Giving the Subordinate Character Covert Agency

Here's the core insight: covert agency is the key. You're not asking the AI to write a character who refuses or fights back openly. You're asking it to write a character who is playing the long game, managing their exposure, rationing their honesty.

A few techniques that work reliably:

Assign the subordinate character a hidden goal. Don't just say "she answers his questions." Say "she's trying to get through this conversation without revealing that she's already contacted HR." Give her something to protect. A specific thing, not a vague emotional state.

Name the character's internal tactics explicitly. In your prompt, tell the AI things like: "she deflects personal questions by asking clarifying questions," or "when cornered, he pivots to bureaucratic language — suddenly everything needs to 'go through the proper channels.'" These are behavioral instructions, not just emotional ones, and they give the AI concrete dialogue moves to execute.

Use the phrase "subtext-heavy dialogue." Literally put those words in your prompt. The AI understands the concept and it noticeably shifts the register of what it generates — characters start speaking around things rather than at them.

Describe what NEITHER character will say directly. "Neither of them will mention that he kissed her at the company party last December. But both of them know they're both thinking about it." This creates a shared secret that pressurizes everything both characters say, even when they're talking about quarterly reports.

Sample Prompts: Boss/Employee, Captor/Captive, Parent/Adult Child

Here are three worked examples. Copy them, adapt them — the structure matters more than the specifics.

Boss / Employee

Why it works: You've given Dana a specific tactic (deflect through process-speak), a clear hidden goal (protect herself without accusing him), and a constraint (no proof yet). The "no action beats" instruction forces the tension to live entirely in the words. Tweak: if you want the scene to escalate, add "by exchange 12, Marcus has dropped the professional courtesy and is asking direct questions she can no longer sidestep."

Captor / Captive

Why it works: This prompt flips the apparent dynamic — Soledad has less structural power but more strategic intelligence in the scene. The AI is being told explicitly that she's performing an emotional state she doesn't have, which prevents the "she nodded meekly but inside she burned" problem. The specific information she's trying to extract (is he alone, how long, does anyone know) gives her dialogue a precise shape. Tweak: add "Reyes says one thing that's accidentally useful to her escape plan" to plant a detail you can pay off later.

Parent / Adult Child

Why it works: Both characters have sophisticated tactics, which means neither one is passive. The phrase "two people playing chess while insisting they're just having a chat" is doing a lot of work — it gives the AI a tonal target. The detail about the late father gives Vivienne a specific weapon that's deniable (she's not threatening, she's just mentioning Dad). Tweak: you could add "at one point Daniel almost breaks, starts to defend himself, then catches himself" to add a crack in his composure.

Iterating the Exchange: How to Escalate or Redistribute Power Mid-Scene

Getting a solid first draft is one thing. The real work is shaping the scene over multiple passes — adjusting the power balance as the conversation unfolds, which is what real scenes do. Power doesn't stay static. Someone says the wrong thing and suddenly the other person has leverage they didn't have ten seconds ago.

The most useful iteration prompt I've found is what I call the pivot moment prompt. You take the draft the AI gave you, identify the exchange where the power should shift, and then write something like:

This is more effective than asking the AI to "make the ending stronger" because it gives a specific mechanical trigger for the change.

Another iteration technique: ask for the scene rewritten from the dominant character's perspective — not to use this version, but to understand what the dominant character thinks is happening. Sometimes the AI will reveal that the dominant character has misread the subordinate completely, which can give you a much more interesting scene if you work that misreading back into the original draft.

You can also ask the AI to write a version where the subordinate character makes one small mistake — breaks for a second, reveals more than they meant to, and then has to recover. This is often more interesting than perfect stoicism, and it tests the dominant character too: does he exploit it, or does he not even notice? That choice says everything about both of them.

A scene where neither character gets quite what they wanted — but both changed something small about their position — is almost always more honest than a clean victory for either side.

Finally, if the dialogue feels too even-keeled no matter what you do, try adding a time pressure to your prompt. Someone has to leave in five minutes. A phone is about to die. The other person in the house might come back. Time pressure forces both characters to make faster decisions, which means their tactics get sloppier, which means more actually gets said than they intended.

The single most practical thing you can do right now: take a dialogue scene you've already drafted — one where the less powerful character feels flat — and write out that power geometry map before you go back to the AI. Structural, relational, informational. It takes four minutes. Then give the subordinate character one specific hidden goal and one specific behavioral tactic. Those two additions alone will change what you get back more than any other technique in this guide.