Escaping tutorial hell: the blank-file ladder for data learners

Why following along feels like learning

When you type code a tutorial shows you, every decision has already been made: which library, which function, which order. Your brain is checking that the code on your screen matches the code in the video. Real work asks you to make those decisions yourself, from nothing. Ten tutorials build the checking skill ten times over and the decision-making skill hardly at all. That's why moving from tutorial to project feels like hitting a wall: the skill you need is one you haven't practised.

The four rungs

Rung 3 is where most people stop and go back to rung 1 with a new tutorial. Don't. Being stuck at rung 3 is the useful part. Every time you have to look up how groupby takes a list of columns, you're building exactly the memory the tutorial couldn't give you.

One technique, all four rungs

Say a tutorial showed you how to clean a messy city column and an amount stored as text.

Rung 1: you type its code and get its output.

Rung 2: you change the data so amounts contain commas ("1,200") and some cities have trailing spaces, then adapt the code until it handles both.

Rung 3: new notebook, tutorial closed. You write the cleaner as a function and check it with assertions, so you know it works without comparing it to anything:

import pandas as pd

# Synthetic messy export
raw = pd.DataFrame({
    "city": ["Noida", "noida ", "Delhi", "DELHI", None],
    "amount": ["1,200", "800", "", "2,200", "650"],
})

def clean_orders(df):
    out = df.copy()
    out["city"] = out["city"].str.strip().str.title()
    out["amount"] = pd.to_numeric(out["amount"].str.replace(",", ""), errors="coerce")
    return out.dropna(subset=["city", "amount"])

clean = clean_orders(raw)
print(clean)
assert list(clean["city"]) == ["Noida", "Noida", "Delhi"]
assert clean["amount"].sum() == 4200

The assertions matter. At rung 3 you have no tutorial output to compare with, so you have to decide in advance what "correct" means. Writing that down is a skill in itself, and tutorials almost never ask you to do it.

Rung 4: you export something real from your own life, like a bank statement CSV with the account numbers removed or a spreadsheet you keep for a hobby, and clean it with the same function, changing it where your data breaks it. The dropped row with no city in the example above raises a real question here: should a row with a missing city be dropped, or kept as "Unknown"? The tutorial never asked. Your data will.

The weekly build-to-watch ratio

Track two numbers each week: hours spent watching or reading instruction, and hours spent writing code without instructions open. Aim for at least one hour of building for every hour of watching. Early on, 1:1 is fine. By the second month, 2:1 is healthier.

If your ratio stays below 1 for two weeks in a row, stop starting new material until it's back up. A backlog of unwatched videos costs you nothing. A backlog of techniques you've never rebuilt is the actual problem.

Signs you are back in the loop

• You've bought or bookmarked a second course on a topic before finishing the first one's exercises.
• Your GitHub has notebooks that are all copies of tutorial projects with the same datasets.
• You can explain what a function does but not when you'd choose it over another.
• You feel productive every day but can't name one thing you built this week without help.

Any one of these is a signal to spend the next session on rung 3.

Where structure helps

Rungs 3 and 4 go faster when someone reviews your from-scratch work and points out the mistake you'd never have found yourself. That's the main value of live classes with mentor feedback. In the Data Science programme, you can bring rung-3 work to a live class and get it reviewed. For a free alternative, free data learning resources that are genuinely good lists exercise-based material. Use it with the ladder, not instead of it.