understand first.
code second.
A Claude Code setup that turns your IDE into a Data Science dojo. 12 steps. One at a time. The code is not the lesson — the thinking is.
The Socratic Ladder
Claude never jumps to the answer. It climbs — asking, hinting, scaffolding, and only revealing as a last resort. The struggle is the lesson.
"What do you think we should do here? Why?" — Claude never opens with code. It opens with curiosity.
A conceptual pointer toward the answer. Analogies, mental models, things to think about — but still no code.
Partial code with gaps for you to complete. The structure is there. You supply the thinking.
Full code with a line-by-line breakdown. Only reached after genuinely working through every other level.
Everything a self-taught DS student needs.
Not an autocomplete tool. Not a shortcut. A learning environment built to make you actually understand what you're doing — not just get outputs.
Every dataset follows the same path: Understand → Explore → Prepare → Model → Improve. The order never changes.
Asks before it tells. Hints before it solves. You never run code you can't explain back.
Auto-saves to PROGRESS.md after every step. Close your laptop, come back tomorrow, pick up exactly where you left off.
Drop a PROBLEM.md with business context and Sensei teaches within that frame — data dictionary, metric, everything.
Reads your data, detects fintech / health / gaming / ecommerce, and adapts every example and analogy accordingly.
Stuck? Ask → nudge → skeleton → reveal. Never jumps straight to the answer — makes you work for it.
Every notebook cell needs a WHAT and WHY comment. Builds intentional habits instead of running cells blindly.
Can't move to the next step until you prove you understood the current one. Learning over output, every time.
12 Steps. One at a Time.
The order never changes. Steps shrink or expand based on your data — but the sequence is fixed, because the thinking process behind DS is universal.
-
01
Dataset OrientationWhat is this data? What's the business question?
-
02
Problem IdentificationClassification? Regression? What's the target?
-
03
Data ProfilingStructure (shape, dtypes, columns) + quality (nulls, duplicates, issues)
-
04
EDADistributions, relationships, outliers
-
05
Train-Test SplitAlways before any transformation
-
06
Feature EngineeringNew features, transformations, domain signals
-
07
Preprocessing PipelineImputation, encoding, scaling — fit on train only
-
08
Baseline ModelSimplest model first. Establish a benchmark.
-
09
Model SelectionTry 2–3 models. Compare with cross-validation.
-
10
EvaluationRight metrics for the right problem.
-
11
Hyperparameter TuningGrid or random search. Understand what changes.
-
12
Final InsightsFeature importance, business interpretation.
6 Commands. Everything You Need.
Type these in Claude Code. That's it — no configuration, no plugins. Just open your project and start learning.
Run this once in any folder to turn it into a Sensei classroom. Writes CLAUDE.md to the folder for future sessions and primes the current session immediately. Re-run it anytime — older classrooms auto-upgrade to the latest ruleset. Then run /start.
The entry point for every session. Reads PROBLEM.md, scans for data, shows your 12-step roadmap, and resumes right where you left off if PROGRESS.md exists — including flagging topics that tripped you up last time so they get revisited.
Gate-checks your understanding with one targeted question. If you answer well, moves forward; missed answers get logged as weak spots to revisit later. Auto-saves PROGRESS.md silently every time — and after Step 12, runs a graduation retrospective.
3 targeted questions for your current step — one at a time, matched to your problem type. One question aims at a logged weak spot; nail it and the weak spot is cleared as mastered. Own the concept, not just the output.
Visual progress tracker showing every step's status. ✅ done, 🔄 in progress, ⬜ not started. Sourced from PROGRESS.md, so it's accurate even across sessions. Shows exactly where you are and what comes next.
Climbs the Socratic Ladder. Ask → nudge → skeleton → reveal. Never jumps straight to the answer. Ends by asking you to explain what you just learned.
Adapts to Any Dataset.
Sensei reads your column names and values, detects the domain, and tailors every example and analogy to match. No generic explanations. No forced analogies.
- Fraud detection
- Credit risk scoring
- Loan eligibility
- Transaction anomalies
- Disease prediction
- Readmission risk
- Lab value analysis
- Diagnosis classification
- Customer churn
- Recommendation systems
- Customer LTV
- Cart abandonment
- Player churn
- Whale detection
- Session analysis
- Level progression
Two Ways to Install.
Pick whichever fits your workflow. Both take under two minutes.