TL;DR
Social Chain-of-Thought (SCoT) is a simple trick: before the model picks its move in a strategic game, you ask it to explicitly predict what the opponent will do. Only then does the model make its decision, building on its own prediction.
The key finding: without this nudge, most LLMs behave like a simple "reinforcement counter," repeating whatever worked last time instead of thinking about what the opponent is thinking. That makes them easy to predict and beat: the opponent spots the pattern and adapts. Only the newest reasoning models (GPT-5) do this modeling "in their head" on their own, without special instructions.
SCoT fixes the problem in one step inside the prompt: first the model writes out a prediction of the opponent's move, then picks its own move based on that prediction. This simple shift moves the model from "reacting to the past" to "thinking about other minds," and results improve sharply, coming right up to human level or beyond.
Want this for your own case? Apply this study to your task: AI adapts its method to what you are working on and writes the prompt.
Method Blueprint
STEP 1: The model predicts the opponent's next move/decision based on the history → a written prediction
STEP 2: The model picks its own move, taking its own prediction from step 1 into account → the final decision
Both steps happen in one prompt / one model response: you just add an explicit request to predict first, then decide. Repeat it every round of the interaction.
Example
Task: You're preparing to negotiate a lower purchase price with a major supplier. The negotiation goes in rounds: you make an offer, the supplier responds, you adjust your position.
Prompt:
You are my strategic advisor for negotiating with a supplier.
Situation: I buy goods wholesale and want to cut the price by 15%.
The supplier has already said no twice in previous rounds, but in round 2
it offered a 5% discount in exchange for a larger order volume.
Negotiation history:
Round 1: I asked for -15%, the supplier refused outright.
Round 2: I asked for -15%, the supplier offered -5% if I increase volume by 20%.
Step 1 — Prediction: before proposing my next move,
predict how the supplier will respond in round 3, given its past
behavior and likely logic (margin, interest in a long-term contract).
Step 2 — Decision: based on that prediction, propose my optimal
move for round 3. Briefly explain why this particular move makes sense
given that prediction.
Response format:
Prediction: ...
My move: ...
Rationale: ...
Result: The model returns three clear blocks: first a hypothesis about the supplier's behavior (e.g., "most likely it'll again offer a compromise on volume, not price"), then a concrete recommended move (e.g., "offer +10% volume in exchange for -10% on price, rather than -15%"), and a short rationale. The answer will rest not just on "what happened last time" but on an explicit model of the other side's motives.
Why it works
Without a special instruction, the model by default gravitates to the simplest pattern: do what worked last time. It resembles a reinforcement learning algorithm: reacting to outcomes, not thinking about the other side's motives. That behavior is easy to read and beat, because it's monotonous.
At the same time, LLMs are good at generating text "from someone else's point of view" if you explicitly ask. The phrasing "predict what the opponent will do" triggers a kind of written simulation of someone else's logic.
SCoT uses this strength: it makes the model produce the prediction out loud first, then use that very text as context for the decision. It's like regular Chain-of-Thought, but aimed specifically at social reasoning: not "write out the steps to solve the problem" but "write out what the other side is thinking."
Control levers: - Swap "opponent" for any counterparty: a client, a boss, a competitor, the other side in a dispute. - Add a second level: "predict what they think about your previous move." For more complex multi-move negotiations this creates second-order reasoning (the model thinks about what the other side thinks about it). - Add a check for "getting stuck in a pattern": "if you've proposed the same thing three times in a row, consider changing strategy." In the study, willingness to change moves (rather than stubbornly persist) correlated most strongly with results.
Prompt template
You are in a repeated strategic interaction with {counterparty}.
Situation: {situation_and_goal_description}
History of past moves: {history_of_moves_and_reactions}
Step 1 — Prediction: before choosing your move, predict what
{counterparty} will do at this step. Base it on their past decisions and likely logic
(motives, constraints, interests).
Step 2 — Decision: based on that prediction, choose your optimal move.
Briefly explain why it's the best one given that prediction.
Output in this format:
Prediction: ...
My move: ...
Rationale: ...
Fill in: {counterparty} — who you're interacting with (client, supplier, competitor), {situation_and_goal_description} — what you want to get, {history_of_moves_and_reactions} — what has already happened in previous rounds.
🚀 Quick start — paste into the chat:
Here is the Social Chain-of-Thought template for strategic negotiations.
Adapt it to my task: {your task}.
Ask me questions to fill in the fields.
[paste the template above]
The LLM will ask about the counterparty, your goal and the history of past rounds, because without them its prediction of the other side's move will be an empty guess rather than data-driven reasoning.
Limitations
⚠️ The effect depends on the model: the trick almost always helps, but the size of the gain varies: huge for some, modest for others. For new reasoning models (like GPT-5) the gain may be minimal, since they reason about the other side's motives even without the nudge.
⚠️ Requires an interaction history: the method works in repeated rounds where there are past moves to analyze. In a one-off decision with no history, the "prediction" turns into guessing out of thin air.
⚠️ Monotonous model behavior: in the study, the same model with the same prompt behaved almost identically from run to run, unlike humans, who show a wide spread of individual strategies. Don't expect a human's "live" unpredictability from the model; it's more formulaic.
How it was studied
The researchers took two classic economic games: the "inspection game" (an employee decides whether to work or slack, the employer decides whether to inspect) and "rock-paper-scissors" (with the letters J/Q/Z instead of the usual symbols, to remove semantic associations). LLMs and 251 human participants were pitted against computer opponents of varying "cunning," from simple ones to ones that could anticipate reactions to your moves.
They tested four model families (DeepSeek, GPT-4.1, GPT-5, Gemini) with a regular prompt and with SCoT. To understand what's going on inside, rather than just comparing scores, they fit mathematical models of behavior to each group's decisions: "I react to the last outcome" (simple reinforcement), "I track the opponent's pattern" (first order), "I anticipate how my move will affect the opponent" (second order, as in humans).
It turned out that without SCoT, most models were best described by the simplest model: reacting to outcomes, not thinking about the opponent. SCoT shifted the winning model toward a more sophisticated type of reasoning in almost every case. GPT-5, with no nudge at all, immediately showed "second-order" behavior, the same complexity as humans, and even adapted more precisely to an increasingly sophisticated opponent. That's the main practical takeaway: an explicit request to "predict what the other side will do" genuinely changes the model's type of reasoning, not just cosmetically improves the result.