No. The specific wrong move is this: you collect a big sample (n > 30), run a Shapiro-Wilk test or eyeball a histogram of the raw values, see they're still skewed, and conclude either that your sample is too small or that the central limit theorem "hasn't kicked in" yet — so you go collect more. It's a tempting move because "n > 30 makes things normal" sounds like it's about the numbers in front of you. It isn't.
Collecting more data does not make your data normal. If your 200 observations are right-skewed, your 5,000 observations will be right-skewed too — usually a sharper skew, because the extra data resolves the shape more clearly. The central limit theorem is not a statement about your data at all. It is a statement about the distribution of the sample mean when you imagine repeating your study many times. Those are two different distributions, and a normality test run on your raw values is checking the wrong one.
So when that skew refuses to go away, you didn't do anything wrong and your sample isn't "too small." You measured the population faithfully, and it happens to be skewed.
Why "n > 30 makes things normal" feels true
Almost every intro course chants some version of "once n is bigger than 30, things become normal." The sentence is real; it just gets cut off before the important noun. The full version is "the distribution of the sample mean becomes approximately normal." Drop those four words and "large sample" and "normal" weld into a single rule that sounds like it applies to the numbers in front of you.
It also matches a reasonable instinct. More data should "fill in" the picture, and somewhere in the back of your mind that filling-in turns into smoothing-out, and smoothing-out turns into a bell. The trouble is that the thing getting smoother is the outline of the true population shape, and if that shape is an exponential decay or a long income tail, a clearer outline is a clearer skew.
The two distributions are almost never drawn side by side for a beginner, so they collapse into one. Separate them and the confusion dissolves.
Two pictures from the same population
Let the population be exponential with rate 1. Mean is 1, it's heavily right-skewed: a tall pile near zero and a thin tail stretching right. Now do two completely different things with it.
Picture 1 — your data. Draw n = 10,000 individual values and histogram them. You get a sharp right-skewed exponential: dense near zero, decaying tail to the right. The skew did not vanish. Compared to a histogram of only 100 values, this one is cleaner: the exponential curve is unmistakable. Ten thousand draws gave you a crisp portrait of a skewed population. This is the picture a normality test sees, and it correctly reports "not normal."
Picture 2 — the sample mean. Now don't plot raw values. Instead, draw a sample of size n = 30, average it, and write down that one number. Repeat 10,000 times. The means come out clustered: 0.95, 1.07, 0.88, 1.12, 1.03, 0.91, and so on. Histogram those 10,000 means and you get a tidy, near-symmetric bell centered on 1, with spread SD of the mean = 1 / sqrt(30) = 0.18.
Same exponential population. Two histograms that look nothing alike. One is the skewed data; the other is the sampling distribution of the mean, and the CLT is a promise about the second one only.
What "n > 30" actually controls
The 30 governs how fast Picture 2 turns into a bell, not whether Picture 1 ever does.
Run Picture 2 again with sample size n = 2: average just two exponential draws at a time, 10,000 times, and histogram the means. You get something visibly lopsided, still leaning right with a clear residual skew. Two values aren't enough to wash out the tail.
Bump the sample size to n = 30 and the same procedure produces a clean bell. That is the whole content of "n > 30": with sample means built from 30 observations each, the sampling distribution is close enough to normal for the usual approximations to hold. Heavier skew needs a larger n; a population that's already roughly symmetric gets there with n far below 30. Thirty is a rule of thumb for the mean's distribution, not a threshold your data crosses.
Notice what changed and what didn't. The population never moved: still exponential, mean 1. The raw-data histogram never moved either. Only the spread and shape of the means changed, tightening from 1/sqrt(2) = 0.71 at n = 2 to 1/sqrt(30) = 0.18 at n = 30 and straightening into a bell. That tightening is its own commonly-confused idea — the spread of the mean (the standard error) shrinks with n while the spread of the data does not, a split worth getting straight on its own in why the standard error shrinks but the standard deviation doesn't.
So what should you check?
If you're deciding whether to trust a t-test or a confidence interval for the mean, you don't need your raw data to be normal. You need the sampling distribution of the mean to be approximately normal, and the CLT hands you that once n is reasonably large and the skew isn't extreme. Checking your 5,000 observations for normality and panicking when they fail is testing a condition the procedure never required. (That confidence interval is a statement about the mean across repeated samples too, which is its own much-misread idea — see what the "95%" in a confidence interval actually refers to.)
If your real goal is to model the individual values, predict a single customer's spend, or set a threshold on one measurement, then the shape of the data matters directly, and no amount of extra sampling will normalize it. You'd fit the skewed distribution itself, or transform it.
The everyday version of this fills stats help forums: someone collects thousands of reaction times or incomes, sees a stubborn right tail, and asks whether they need an even bigger sample to "trigger" the CLT. Nothing they collect will move that tail, and nothing about that tail blocks the CLT from doing its job on the mean.
Check yourself
A researcher has 8,000 household incomes. The histogram is sharply right-skewed. She wants a confidence interval for the mean household income. Which statement is correct?
A) She must collect more data until the income histogram looks normal. B) The 8,000 raw incomes will stay skewed; the CLT makes the sampling distribution of the mean approximately normal, so the interval is fine. C) Skewed raw data means the CLT does not apply and the mean is meaningless. D) Running a normality test on the 8,000 incomes tells her whether the CLT has "kicked in."
Correct answer: B.
The raw incomes describe the population and will stay skewed no matter how many she collects — more data just sharpens the skew. The CLT operates on the distribution of the sample mean across hypothetical repeated samples, and with n = 8,000 that distribution is very close to normal, so a standard interval for the mean is justified. A confuses data shape with the mean's shape. C invents a restriction the CLT doesn't have. D tests the wrong distribution entirely.
Close the gap
The fix here isn't a harder formula — it's keeping two pictures separate that textbooks routinely overlay: the distribution of your data and the distribution of its mean. Once a learner can say which one a question is about, "is my sample big enough to be normal" stops being a question and the right tool becomes obvious. Tracking exactly where that split breaks down, and re-anchoring it the moment it does, is what Gradual Learning is built to do.