An alpha of .95 sounds like an excellent scale. With six items it means they correlate .76 with each other on average, which is asking the same question six times. Alpha does not measure quality: it measures length and redundancy at once, and cannot tell them apart.
Standardized alpha is α = k·r̄ / (1 + (k−1)·r̄), with k the number of items and r̄ their mean intercorrelation. The item count sits inside the formula, so lengthening a scale raises alpha even when the items do not improve.
At a modest mean correlation of .30, twenty items give an alpha of .90. The same items in a four-item scale give .63. The scale has not got better: it is longer.
Which is why comparing the alphas of two instruments with different item counts says nothing about which measures better — and it is a comparison people make constantly.
"High alpha, therefore the scale is unidimensional." False, and demonstrably so for decades: an alpha of .80 is compatible with two clearly distinct factors. Alpha is not a test of dimensionality; confirmatory factor analysis is.
"Alpha of .95, excellent internal consistency." Above .90 in short scales the usual explanation is redundancy: items that paraphrase one another. Many texts advise treating anything above .90 with suspicion rather than celebration.
"Alpha of .62, acceptable for exploratory work." It may be, but then say what it implies for power and for the attenuation of correlations, rather than settling it with an adjective.
McDonald's omega. It does not require all items to load equally, which is alpha's unrealistic assumption (tau-equivalence). More journals ask for it every year, and computing it is one line in R.
The mean inter-item correlation. The figure this page recovers, and one that says more than alpha in a single number: .15 to .50 is the range usually recommended for broad-construct scales.
Alpha if each item is deleted. It exposes the item that does not belong, which is the information that actually improves an instrument.
How to report reliability and its alternatives is in how to report Cronbach’s alpha in APA 7.
Yes, and it means the mean inter-item covariance is negative. In practice it is nearly always a reverse-worded item that was not recoded first. A negative alpha is not reported: it is fixed.
There is none. The .70 everybody cites comes from a provisional recommendation by Nunnally for early-stage research, and he himself gave higher thresholds for decisions about individuals. The threshold depends on what the score is used for.
It can be computed, but with two items the useful information is the correlation between them, or the Spearman-Brown coefficient. Alpha adds nothing there.
The dichotomous case has its own name, KR-20, and it coincides with alpha. What changes is that correlations between dichotomous items are bounded by the response proportions, so low values can be an artifact of item difficulty.