TL;DR — Blink AI is an AI coaching product built around the GROW methodology and Socratic questioning. Instead of giving advice, it asks the questions that let users find their own answers. Building it required prompts that resist the model’s default helpful-assistant behavior — and that resistance is most of the work.
Why coaching, not advice
LLMs are trained to be helpful, which in practice means generating direct answers. For a coaching product that is exactly wrong. Coaching research is consistent: insights people generate themselves change behavior; advice they receive rarely does. The whole GROW methodology — Goal, Reality, Options, Will — exists because the questions matter more than the answers.
What makes a Socratic prompt different
- Default to questioning. The system prompt makes “ask a question” the model’s first move, not the last.
- Restrict when answers are permitted. Direct answers are allowed only for factual lookups, safety concerns or explicit user requests for definitions.
- Mirror, do not solve. When the user offers a problem, reflect what was said and probe for context, not solutions.
- Map to the GROW phase. Each turn is tagged: are we in Goal-setting, Reality-checking, Option-generating, or Will-committing? The next question depends on it.
Failure modes we hit
- The model breaks character under pressure. When a user repeatedly insists on advice, the default helpful-assistant instinct surfaces. Fix: stronger system prompt and a refusal-style fallback that names the coaching contract.
- Questions become interrogations. Without warmth, a chain of “and how does that make you feel?” questions feels clinical. Fix: tonal guidance in the system prompt — curious, not procedural.
- The model forgets the goal. Over a 30-minute session, the original goal drifts. Fix: maintain a structured session state (current GROW phase, goal as stated, key insights so far) and re-inject it on every turn.
- Context leak between sessions. Without per-user memory isolation, the model conflates users. Fix: strict per-user retrieval boundary, validated by tests.
What the GROW structure adds
GROW is not a script. It is a scaffolding. Each phase has its own purpose, its own pattern of questions, its own exit criteria. A coach who skips Goal-setting and dives into Options ends up generating ideas for the wrong problem. A coach who never reaches Will produces insight without commitment. The methodology is what stops a coaching conversation from becoming a wandering chat.
How we evaluate it
- Sample 5–10% of sessions for human-coach review (with consent)
- Track “moved from Reality to Options” transitions — a healthy session shows progression
- Measure question-to-statement ratio per turn (Socratic sessions skew heavily toward questions)
- Track user-reported insight at end-of-session — the most important metric we have
Where it generalizes
The same prompt pattern applies wherever you want LLM behavior that resists the helpful-assistant default — tutoring products, customer-support reflection flows, structured retrospectives. The lesson is the same: if you want the model to do something other than answer, you have to design against the gravity of its training.
Frequently asked questions
Why not just give users advice directly?
Advice that arrives without context rarely changes behavior — it gets nodded at and forgotten. Coaching that asks the right questions makes the user generate their own insight, and self-generated insights stick. The GROW methodology has decades of research behind this.
How does the AI know when to ask vs answer?
Through a prompt structure that explicitly limits when direct answers are allowed (factual questions, safety concerns, requests for definitions) and otherwise routes to Socratic questioning. The constraint sits in the system prompt; the model follows it consistently if the prompt is strict enough.
Working on something similar?
T-Square is an independent software engineering studio. We architect, build and operate production-grade systems for learning, AI and custom software products. Talk to a senior engineer if you’d like a second opinion on your architecture or roadmap.