The Center for the Advancement of Population Assessment Methodology (CAPAM) hosted a technical workshop on Recruitment: theory, estimation, and application in fishery stock assessment models in Miami, Florida, USA, from 30 October to 3 November 2017. The recruitment workshop was the fourth in a series organized by CAPAM as part of its Good Practices in Stock Assessment Modeling Program for improving fishery stock assessments. The workshop was sponsored by NOAA and the University of Miami via the Cooperative Institute for Marine and Atmospheric Science (CIMAS). The primary goal of the workshop was to provide advice and guidance on practices for modeling recruitment in fishery assessments. The focus was on model specification, parameter estimation, and management consequences. The five-day forum included an interactive modeling session, six keynote addresses, and 30 research presentations; discussions focused on major topics, from describing mechanistic properties of recruitment to time series modeling and management implications. Ninety-five attendees registered, and an average of 20 people were online at any given time. A special issue in the journal Fisheries Research finalized in March 2019 features articles from the workshop. This report summarizes the presentations and discussions made during the workshop. As such, it represents the general views of the editors, rather than any achieved consensus set of recommendations. Several important research topics on recruitment are identified to guide further research, along with recommended practices to consider when developing stock assessment models. Some of the key recommendations discussed at the workshop were: (1) It is important to have good fishery-independent recruitment information, not only for improving assessment models, but also as the basis for an early warning sign. A series of low recruitments is often the first indication that a stock is in trouble. Waiting until the signal is observed in the catch data (if the data are good enough to show it) can be too late. (2) Each assessment should describe the recruitment process, and evaluate alternative hypotheses and what they imply about the stock–recruitment relationship. (3) In cases where the parameters of the stock–recruitment relationship are considered estimable, log-likelihood profiles and other diagnostic tools should be examined to determine how reliable the estimates are. A useful diagnostic, when possible, is to compare the average recruitment (or bias-adjusted deviates, if using a stock–recruitment relationship) over a period where recruitment is considered likely to fluctuate about R0. (4) Random effects models should improve the estimability of the parameters of the stock–recruitment relationship and ?r in principle, but will not resolve the problem if the data are not informative. (5) Include and estimate autocorrelation about the stock–recruitment relationship within the assessment. Autocorrelation reduces short-term uncertainty, but increases uncertainty in long-term projections. (6) The effects of assuming alternative stock–recruitment models should be evaluated, unless there is clear reason to expect that the chosen model is correct. (7) In many cases, it will be more cost-effective to develop management procedures and harvest control rules that are robust to recruitment uncertainty, rather than attempting to incorporate that uncertainty into assessment models. (8) Continued research on environmental drivers and spatial dynamics influencing recruitment should be encouraged.