Multiple-day diet records can be unsuitable for cohort studies because of high administrative and analytical costs. Costs could be reduced if a subsample of participants were analyzed in a nested case-control study. However, completed records are usually reviewed (‘documented’) with participants to correct errors and omissions before analysis. The authors evaluated the suitability of using undocumented 3-day food records in 2 samples of women in a Northern California cohort study of breast cancer survivorship (2006-2009). One group of participants (n = 130) received an introduction to the food record at enrollment, while another (n = 70) received more comprehensive instruction. Food records were mailed to participants 6 months later for follow-up and were analyzed as received and after phone documentation. Error rates for adequate completion were high in the first group but substantially lower among persons receiving instruction; prevalences of missing data on serving size and incomplete food descriptions changed from 30% to 4% and from 32% to 6%, respectively (P < 0.0001). Correlations between nutrient intakes calculated from undocumented and documented records were 0.72-0.93 in the first group and were significantly stronger (0.84-0.99) among persons receiving instruction. Documentation had little effect on intraclass correlation coefficients across days, but training increased the coefficients for many nutrients. When participants receive proper instruction, undocumented food records can be satisfactory for large epidemiologic studies.