How do managers affect worker productivity and retention through feedback? Using data from GitHub and LinkedIn, I analyze over 230 million feedback messages from code reviews across 1.7 million software teams. I use large language models to classify feedback tone (toxicity, positivity) and information (constructiveness). Exploiting the random assignment of code reviewers, I find that toxic feedback reduces workers' subsequent code quantity and quality, whereas nontoxic criticism has no comparable effect. Positive feedback raises productivity. These effects extend beyond code output: positive feedback also raises firm retention within a year, and effects spill over to coworkers. Information content shifts workers toward revising existing tasks and away from new development. Linking these effects back to managers, I find that feedback explains 22% of the variation in manager quality, measured as value added to worker productivity. This paper shows that manager feedback style is not only a workplace amenity, but also a factor that affects worker productivity and a measurable component of manager quality.

Does the division of labor increase team productivity? This paper provides new evidence challenging the conventional view that specialization increases productivity. I create a panel dataset from GitHub, covering 35 million task allocations across 64,400 software development teams from 2017 to 2023. My result shows a negative relationship between team specialization and various productivity metrics, including output quality, quantity, and user issue resolution time. To identify causal effects, I exploit GitHub’s introduction of an automatic task assignment feature, which evenly distributes tasks across team members. Using a matched difference-in-differences design, I find that adoption of this feature reduces specialization and leads to significant gains in productivity: output quality rises by 4%, output quantity by 21%. Team communication also increases by 13%, suggesting that improved interaction and knowledge exchange are a key mechanism behind these productivity gains. These findings highlight a trade-off in non-routine production: while specialization increases task-specific human capital, it impedes cross-task knowledge spillovers that are essential for innovation.