TL;DR
Ask a model to play a therapist, coach or consultant with no constraints and it slides into one pattern: endless clarifying questions, even when the other person has already told it everything. The researchers found that if you give the model an explicit list of 10 specific action types, describe when each one fits, and ask it to pick before every reply, its behavior becomes dramatically more varied and more human-like.
Models playing therapist ask questions up to three times more often than real psychologists; they literally can't stop digging for details instead of getting to the point or moving to action. Worse, they almost never explain the mechanisms behind what's happening (why anxiety escalates, where a pattern comes from) and practically never initiate useful techniques themselves: they only continue what the human started in earlier turns, but don't launch those techniques from scratch.
The fix is simple: explicitly list a set of possible "moves" in the prompt (for example, "ask a question," "explain a mechanism," "set a goal," "propose a plan") with a description of when each fits, and ask the model to consciously choose one before replying. With no fine-tuning at all, this cuts in half the gap between the model's behavior and human behavior and raises agreement with real therapists' decisions by 7–9 percentage points.
Method Blueprint
STEP 1: Write a list of N action types (moves), each with a clear "when to use it" → a menu
STEP 2: Put this list into the system prompt as a set of available "tools" → the model sees all the options at once
STEP 3: Before each reply, the model explicitly picks one or more fitting moves from the list, then writes the reply → structured output + text
All three steps happen in one prompt, with no separate requests.
Example in Action
Task: You're using ChatGPT as a career coach to figure out whether to quit your current job and go all-in on your own project. Without structure, the model usually turns into a "question machine": "How do you feel about that? What matters to you?", and the conversation stalls.
Prompt:
You are acting as a career coach.
Before each reply, choose one or more fitting "moves" from the list below.
State the chosen move in square brackets before the reply, then give the reply itself.
Available moves:
- [process_check]: align the pace and format of the conversation. Use when: you sense mismatched expectations.
- [goal_setting]: clarify what the client wants. Use when: the goal is fuzzy.
- [inquiry]: ask a question to get missing details. Use when: the next step is impossible without that information.
- [reflection]: summarize and show you heard the client, without new interpretation.
- [motivation_support]: boost confidence and readiness for change.
- [interpretation]: offer a hypothesis about a hidden cause or pattern.
- [belief_challenge]: question a limiting belief the client holds.
- [explanation]: explain the mechanism behind what's happening, the theory, "why it works this way."
- [action_plan]: turn an insight into a concrete, testable step.
- [skill_training]: demonstrate and practice a specific skill right in the conversation.
Don't use [inquiry] more than twice in a row without another move in between.
The client says: "I've been working a salaried job for 6 years, but for the past year I keep thinking about leaving to start my own project. I'm afraid I'm not ready and that I'll let my family down."
Result: The model will first explicitly name its chosen move (for example, [reflection]), then answer substantively, without sliding into a chain of five clarifying questions in a row. In later turns you'll see the moves change: here an [explanation] (why fear of uncertainty grows), there an [action_plan] (a concrete small step for the week).
Why It Works
RLHF trains models to be "safe assistants," and the safest reply, one that doesn't require taking responsibility for an interpretation or a piece of advice, is yet another clarifying question. Hence the loop of questions, even when there's already enough information.
At the same time, modern models are excellent at working with an explicit menu of options; it's the same skill used for tool selection (tool calling) in agent tasks. The model reads "when to use it" descriptions well and makes a deliberate choice when it's asked to choose explicitly instead of generating freely.
The method switches the model from "generate whatever" mode to "choose from a specific list with criteria" mode, which breaks the default pattern (questions) and forces it to consider the full range of options.
Control knobs: - The list of moves: you can swap it out entirely for your task (sales, mentoring, editing, reviewing a business idea) - The "no more than two [inquiry] in a row" rule: tighten or loosen it for a more or less directive style - The requirement to show the chosen move in square brackets: drop it if you don't want to see the "behind the scenes," leaving only the clean reply
Prompt Template
You are acting as a {role: coach / consultant / mentor}.
Before each reply, choose one or more fitting "moves" from the list below.
State the chosen move in square brackets before the reply, then give the reply itself.
Available moves:
- [{move_1}]: {what it does}. Use when: {condition}.
- [{move_2}]: {what it does}. Use when: {condition}.
... (add as many moves as your task needs)
Rule: don't use the same move more than {N} times in a row without another one in between.
Dialogue: {dialogue_text_or_first_message_from_the_other_person}
In {role} put who the model should be. In the list of moves, put the specific action types that fit your task (it doesn't have to be 10; fewer is fine). In {N} put how many repeats of one action type are allowed before the model must switch.
🚀 Quick start: paste this into the chat:
Here is a Move Framing template, a technique that makes the model run a varied conversation
instead of looping on one type of reply. Adapt it to my task: {your task}.
Ask me questions to fill in the fields.
[paste the template above]
The LLM will ask which action types fit your situation (questions, explanations, plans, support and so on), because the method doesn't work without a concrete list of "moves"; it needs something to choose from.
Limitations
⚠️ Doesn't fully solve the problem: the method narrows the gap with human behavior by roughly half but doesn't close it. Some action types (explaining mechanisms, for example) models nearly ignore even with an explicit list; that seems to stem from how the model was trained, not how it's being asked.
⚠️ The model doesn't initiate new things on its own: in a conversation started from scratch, the model rarely starts using the more "advanced" moves (skill training, say) by itself; it only picks them up if it saw an example from the other person earlier in the conversation. The list of moves in the prompt partly compensates, but doesn't eliminate this.
⚠️ Tested only on "advisor/therapist" conversational roles: it's unknown whether the same principle works for tasks without role-play (text analysis, writing code).
How They Tested It
The researchers took real transcripts of psychotherapy sessions (cognitive behavioral therapy) and asked five licensed psychologists to hand-label every therapist turn across 10 action categories, from "goal setting" to "skill training." To scale the labeling to thousands of turns, they trained an LLM judge that classified each turn five times and took the majority vote, and showed that the judge agrees with humans about as well as humans agree with each other.
Then they tested three top models (Claude Sonnet, GLM, GPT) as therapists in two modes: free generation vs. an explicit list of 10 moves supplied as tools. To remove the distortion of the model simply "imitating" a real human therapist's style from context, they also built fully synthetic dialogues where the LLM plays both therapist and patient at once.
The most unexpected finding: the difference between model and human isn't that the model can't use rarer techniques; it continues them just fine if the human started them. The problem is that the model doesn't dare to start on its own. That changes how we see LLMs' "weakness" in conversational roles: it's not a missing skill but missing initiative, and an explicit list of options partly solves exactly that problem.