We conduct a two-part study to better understand pen grip postures for general input like mode switching and command invocation. The first part of the study asks participants what variations of their normal pen grip posture they might use, without any specific consideration for sensing capabilities. The second part evaluates three of their suggested postures with an additional set of six postures designed for the sensing capabilities of a consumer EMG armband. Results show that grips considered normal and mature, such as the dynamic tripod and the dynamic quadrupod, are the best candidates for pen-grip based interaction, followed by finger-on-pen postures and grips using pen tilt. A convolutional neural network trained on EMG data gathered during the study yields above 70% within-participant recognition accuracy for common sets of five postures and above 80% for three-posture subsets. Based on the results, we propose design guidelines for pen interaction using variations of grip postures.