AI does the drafting. You do the thinking. Five moves that turn AI from a shortcut into real, durable skill — the difference between an answer you borrowed and one you can explain, defend, and recreate.
01
Frame with purpose and constraints.
“Know what "great" looks like — and what the limits are — before you touch AI.”
Most students open with "what do I do?" You open with what a high-quality result even is, where the data or problem breaks down, and what risks — practical or ethical — are in play.
Example prompt
Act as a senior expert in this field. Given my goal and data, what questions are realistically answerable? What defines a strong result — and how do we avoid misleading conclusions?
WHY
You internalize purpose and learn to recognize uncertainty, instead of chasing a deliverable.
02
Map the reasoning of real experts.
“Break the task into reasoning layers — not code chunks or tasks to hand off.”
Shift from procedural to conceptual decomposition: what decisions does each stage demand, what assumptions must be tested, and what would even count as a non-trivial finding?
Example prompt
Decompose this problem into the stages a top expert would use. For each stage, list the key decisions, the assumptions to test, and what would signal a non-trivial finding.
WHY
You build meta-reasoning — learning not just what to do, but how experts think through ambiguity.
03
Force transparency and thoughtfulness.
“Never take an output without the "why."”
Require the AI to show intermediate steps, explain why it chose one method over another, and flag its own uncertainty — so you interrogate the credibility of every result.
Example prompt
Do the work, but with every result include (1) why you chose this method, (2) how confident you are, and (3) where it might be wrong.
WHY
You build interpretability literacy — you stop accepting outputs at face value.
04
Cross-validate with science, peers, and reality.
“Truth isn't what AI says — it's what survives scrutiny.”
No insight is accepted until it is triangulated: against published work, against the raw ground truth, and against a peer or mentor who can sanity-check it.
Example prompt
For each claim, cite a source that supports or contradicts it, or propose a test — a control, a permutation — that would prove it robust.
WHY
You develop scientific skepticism and move toward publishing-level standards.
05
Stress-test what you actually internalized.
“If you can't rebuild it without AI, you haven't learned it.”
Close the AI and reproduce the core from scratch. Then compare your version to the AI's and annotate every difference in logic or framing — the delta is where the learning lives.
Example prompt
Now I'll close the AI and redo the core myself. I'll document what I remembered, what differed, and what I learned by rebuilding it.
WHY
Active retrieval and reflection turn passive review into durable mastery and skill transfer.
These aren't teaching demos — they're production systems our mentor designs and ships. Students work on real modules of them (named below, first name only). That's the difference between learning from someone who does the work and learning from a slide deck.
One method, four real systems. Four of these builds — targetai, larchai, musicASD, and AuraDiary — run on the same architecture we teach: give the AI a written constitution and MCP tools, then keep a human in the loop. It's not a tutorial — it's how our mentor actually builds.
Flagship
An AI college-admissions platform for counselors — college match, essay coaching, document intelligence, all run by a constitution-governed agent over 6 MCP servers.
Students who contributed
Ping · Xing · Lixin · Paige
A production investing-research platform — data pipelines, backtesting, and LLM agents with a written constitution and MCP tools.
Students who contributed
Jasmine · Lucy · William · Johnny · Alvin · Alan
An AI job-search copilot: discover → score → tailor → prep → land, with a career-vault RAG and a multi-LLM cost router.
Students who contributed
Allen · Leo · Patrick
A clinical music-therapy platform for children with autism — AI drafts, the board-certified therapist signs. Safety-reviewed generation + a Music-DNA profile.
Students who contributed
Victoria · Jerry · Chris
An AI journaling app with an event-driven agent, an evolving persona, and a diary-to-blog publishing pipeline.
Students who contributed
Eric · Anita · Elisa
A research pipeline that trains a neural net to read emotion from biosignals and generate matching music — real model training on real datasets.
Students who contributed
Ethan · York · Wang
A research prototype (in development) that generates emotionally-aligned music for silent video using a custom transformer + LSTM architecture.
Students who contributed
Grant · Jack
An AI debate coach: debate an AI opponent, then get scored and coached against a structured rubric. Used by our 2023 summer-camp cohort.
Students who contributed
2023 Summer Camp cohort
Mentor builds are shown by architecture and UI; the code stays private. Students are credited by first name for the modules they worked on — not as authors of the whole system.
AI thinking with you
= Learning.
AI thinking for you
= No learning.