The technical capability to design interactive and adaptive learning environments is maturing rapidly. In this paper, we report on an empirical investigation that used a variant of the wizard-of-oz methodology to identify guidelines for the design of effective simulated tutors within this type of environment. Thirty-six subjects completed a 4 hour curriculum on simple direct current (D.C.) circuits while interacting with an unseen human tutor via a chat interface. Student learning, calculated using pre-test and post-test scores, were significantly correlated with the amount that the students talked, as measured by the average number of words per student turn. Also, the use of open-ended questions resulted in a statistically significant increase in the length of student utterances. Tutor strategies that were significantly positively predictive of student learning gains included reinforcing correct responses by restating the answer and remediating incorrect responses by proposing activities likely to help the student figure out the correct answer by him or herself. We suggest that the strategies that proved effective for human tutors are good candidates for incorporation into a simulated tutor. The paper ends with a discussion of the implications for the design of simulated tutors, the limitations of the current research, and additional research requirements.