when the mean differs from zero by more than 2 units of (estimated) standard deviation, then we can reject with 95% confidence the null hypothesis that the true population mean equals zero.Now, you might worry about the fact that the CLT explicitly involves taking the limit of the estimator's distribution as the sample size grows without bound (the mathematical term is "diverges", Paul Gowder). To paraphrase Eugene, no sample size ever actually is infinite, so how can we use the CLT?
When the sample size is large enough that a given rejection rule would not be affected by replacing the normal approximation with the true distribution function, then the sample size is almost infinite.This definition, which is analogous to (a special case of?) more general suggestions in the comments above, has one disadvantage: whether or not the sample size is almost infinite will depend on the level of precision chosen for the rejection rule (e.g., do we use 2 standard deviations, 1.96, or even a more precise figure?). In other words, "almost infinite" has a contingent definition. But in practice that won't much matter, especially if there are clear conventions in place. And anyway, lots of mathematical concepts are contingent (for example, a faster convergence rate for an estimator requires that we have a large "enough" sample size to have any bite).