Suggestions from @hadley

• Require drop = FALSE whenever [ is used with two or more args
• Require na.rm argument for any function that has it (to force you to
  think about missing values during development)
• Don't use of require()/library() inside a function
• Don't change global state: setwd(), options(), par() etc. Could
  have next package robustr that contained alternatives (i.e. the with_*
  functions from devtools)
• Always turn partial matching warning options on (not really code
  linting, but related)
• Avoid functions that use NSE (with(), subset(), transform() etc)
• Avoid functions that return different types of output (e.g. sapply())
• TRUE and FALSE instead of T and F.
• return() calls at the end of functions (result of last statement is always returned, so the explicit return() is redundant (and also slower as it requires an additional function call).
• use of 1:length(x) in for loops (dangerous if length(x) can be 0) as 1:0 returns c(1, 0), should use seq_len(length(x)) instead.
• Iterative appending to vectors within loops rather than pre-allocation the vectors and filling. (the former is quadratic in time complexity, as R has to keep re-sizing the vector on every assignment.

# bad
for (i in seq_len(10)) {
  x <- c(x, i)
}

# good
x <- integer(10)
for (i in seq_len(10)) {
  x[i] <- i
}