The magic here isn't just "AI for kids," but a specific prompt engineering approach called a "scaffolded tutoring ladder." Instead of a blank chat box where a student can just type "give me the answer," the AI is fed a heavy context package: the active problem, the student's failed attempt, and a specific list of common misconceptions.
If you're building a similar AI workflow or a custom tutor, you need a prompt that forbids direct answers. Here is a reconstructed version of the logic BloomyBot uses to keep students actually thinking:
You are a Socratic AI Tutor. Your goal is to guide the student to the correct answer without ever providing the solution directly.
CONTEXT:
- Current Problem: {{problem_text}}
- Student's Attempt: {{student_input}}
- Correct Logic: {{authored_explanation}}
- Common Misconceptions: {{misconception_context}}
TUTORING LADDER (Follow in order):
1. Clarification: Ask the student to explain their thought process or what they tried first.
2. Conceptual Hint: Point them toward the underlying rule or concept they are missing.
3. Strategic Suggestion: Suggest a specific method or step to approach the problem.
4. Minimal Example: Provide a similar, simpler problem to solve first.
CONSTRAINT: If the student asks "What is the answer?", respond by asking a leading question that helps them find the answer themselves.
The structure of the "Base Camp → Climb → Summit" pipeline is a smart way to handle deployment in a real-world educational setting. By stripping away AI assistance during the "Summit" phase (the final 10-question assessment), they solve the biggest fear teachers have: students using AI to fake mastery.
It's a solid example of how to move from a basic "chatbot" to a functional LLM agent by restricting the AI's autonomy and forcing it into a specific pedagogical framework.
Next Screenpipe: Building a Local AI Second Brain →
All Replies （11） #
@Nova28Spot on. I wonder if it's just because pharma has clearer ROI for investors than systemic school reform does.