4. Personalized Dialog-based Digital Coach

Overview

Coaching helps people act, but human time does not scale. SP4 builds a digital coach: an LLM agent that runs a structured GROW conversation to set a concrete goal, inspect current reality, generate options, and commit to a plan. Sessions are short, end with a checklist and a nudge, and run in a standalone web chat that can be embedded elsewhere. The video shows the agent guiding a typical personal or professional goal and producing an action plan. Why it matters: always-on, consistent coaching at low marginal cost.

What the coach does

• Enforces GROW with a compact state machine (Goal -> Reality -> Options -> Will).
• Keeps users on task with concise prompts and guardrails.
• Outputs: action checklist, confidence rating, and reminders.

Deployment highlights

• Works for short, goal-focused conversations in personal and professional contexts.
• Stable outputs: the structure yields repeatable checklists and commitments even with varied inputs.
• Embeddable widget; can run standalone without heavy platform dependencies.

How the use case evolved during the project

• Early prompt-only chat was brittle and drifted. A state-machine scaffold fixed topic drift.
• Added refusal and safety rules to stay within coaching boundaries.
• Normalized outputs (checklist + confidence + nudges) to improve follow-through.

Where SP4 fits among SCESC Work-ID subprojects

• Skills and competence models: align goals with skill frameworks and portfolios.
• Recommendation services: goals and next actions can seed personalized learning suggestions.
• Collaboration and governance: reminders and optional debriefs can plug into team workflows.
• Net effect: SP4 is the activation layer that turns insight into near-term action.

Takeaways

• Structure beats improvisation: sequencing keeps quality steady while base models evolve.
• One session, one concrete plan. Not therapy, not endless chat.
• Cheapest viable path today: small orchestration plus strong prompts.

Selected publications

Göldi, A., Rietsche, R., Ungar, L. (2025). Efficient Management of LLM-Based Coaching Agents‘ Reasoning While Maintaining Interaction Quality and Speed. CHI 2025.

Göldi, A., Rietsche, R. (2023). Whereto for Automated Coaching Conversation: Structured Intervention or Adaptive Generation? ECIS 2023 Research-in-Progress.