Species interactions in food webs are usually recognized as dynamic, varying across species, space, and time because of biotic and abiotic drivers. Yet food webs also show emergent properties that appear consistent, such as a skewed frequency distribution of interaction strengths (many weak, few strong). Reconciling these two properties requires an understanding of the variation in pairwise interaction strengths and its underlying mechanisms. We estimated stream sculpin feeding rates in three seasons at nine sites in Oregon to examine variation in trophic interaction strengths both across and within predator–prey pairs. Predator and prey densities, prey body mass, and abiotic factors were considered as putative drivers of within‐pair variation over space and time. We hypothesized that consistently skewed interaction strength distributions could result if individual interaction strengths show relatively little variation, or alternatively, if interaction strengths vary but shift in ways that conserve their overall frequency distribution. Feeding rate distributions remained consistently and positively skewed across all sites and seasons. The mean coefficient of variation in feeding rates within each of 25 focal species pairs across surveys was less than half the mean coefficient of variation seen across species pairs within a survey. The rank order of feeding rates also remained conserved across streams, seasons and individual surveys. On average, feeding rates on each prey taxon nonetheless varied by a hundredfold, with some feeding rates showing more variation in space and others in time. In general, feeding rates increased with prey density and decreased with high stream flows and low water temperatures, although for nearly half of all species pairs, factors other than prey density explained the most variation. Our findings show that although individual interaction strengths exhibit considerable variation in space and time, they can nonetheless remain relatively consistent, and thus predictable, compared to the even larger variation that occurs across species pairs. These results highlight how the ecological scale of inference can strongly shape conclusions about interaction strength consistency and help reconcile how the skewed nature of interaction strength distributions can persist in highly dynamic food webs.