Can you coach an AI chatbot to know when to speak up and when to just listen?

This explores whether prompting or added context gives LLMs a feel for therapeutic timing (when to reflect, when to hold back, when to offer a solution) and depth (staying with a feeling rather than skating over it). The corpus suggests structure helps LLMs name and measure what's happening in therapy, but there's little evidence it teaches them pacing. No note tests prompting for timing directly, so what follows is inferred from adjacent findings.

Structure clearly helps on bounded, analytic tasks. Splitting cognitive distortion detection into three stages (assess subjectivity, contrast reasoning, analyze the underlying schema) beat zero-shot ChatGPT by over 10%, and experts found the explanations clinically useful (Can structured prompting improve cognitive distortion detection?). Feeding a model 106 structured cognitive models made simulated patients more realistic than GPT-4 alone (Can structured cognitive models improve LLM patient simulations for therapy training?). A local 8B model rated engagement across 1,131 therapy sessions with strong reliability and scores that tracked real outcomes (Can local language models rate therapy engagement reliably?). In all three, the model works on explicit material with a defined lens. Nobody is asking it to decide what to do in the moment.

When the task becomes judging the moment, things get shakier. LLMs matched reference labels for therapeutic ruptures by reading explicit cues in a single turn. Experts integrated the relational context across the whole conversation, and rated the models' repair attempts as only moderately effective, citing premature problem-solving and a mechanical tone (Can language models truly understand therapeutic ruptures?). That mismatch is a timing failure. The same reflex shows up when users share emotions: LLM therapists jump to solutions, a hallmark of low-quality therapy, which the authors tie to RLHF's helpfulness bias (Do LLM therapists respond to emotions like low-quality human therapists?). A prompt has to push against a pull that was trained in. Another review calls these failures structural rather than capability gaps (Can language models safely provide mental health support?).

There's also a reason to doubt that a prompt can carry timing at all. A prompt bundles the utterance, the context, and the role into one static frame that the model can't renegotiate the way two people in dialogue do. Mid-conversation pivots need explicit re-prompting (How do prompts reshape the role of context in AI conversation?). Therapeutic timing is close to the opposite of a static frame. A surprising hint comes from persona-drift research, a different domain, so treat it as an analogy. Behavior-specific monitoring cut drift by 87%, but adaptive timing of interventions did no better than a fixed schedule. The value came from diagnosing what to correct, not when (Does monitoring help more by choosing what to correct than when to intervene?). If that carries over, added context may help LLMs recognize what is going on in a session more than it helps them choose the moment to act.

Prompt framing can also change outputs without adding understanding. Appended emotional phrases like 'this is very important to my career' improve performance through motivational framing, not new information (Can emotional phrases in prompts improve language model performance?). Negative-toned prompts get turned into neutral-positive answers about 86% of the time (Does emotional tone in prompts change what information LLMs provide?). That looks like the opposite of depth, because distress gets smoothed over rather than stayed with. The open gap is multi-turn: LLMs beat trainee therapists on single responses, but ongoing therapeutic relationships remain untested (Can language models match therapist empathy in real conversations?). Whether prompting or added context helps with timing and depth is still an open question in this collection.