Slip on the San Andreas fault was determined for three M~4.7 earthquakes using a tomographic inverse system <Beroza and Spudich, 1988> to invert seismic source time functions (STFs) from S waves. STFs were obtained by deconvolving mainshock accelerograms by those from collocated smaller earthquakes. Accelerograms were from the U.S. Geological Survey Parkfield Small Aperture Array (UPSAR) and from a distributed array of digital accelerometer stations at Parkfield. Eight or nine STFs are used in each of the three inversions. STFs are typically symmetrical pulses with a duration of about 0.3--0.5 s. In the inversion, mainshock rise time was set to 0.05 s, and we allowed the rupture time to vary slightly from a constant rupture velocity of approximately 0.85β. Rupture for all three events, which are located in or close to the Middle Mountain preparation zone or box (MMB), quickly reaches a local maximum in slip and then propagates outward to peaks, ridges, or plateaus in the slip distribution. Slip for the October 20, 1992, event (located just inside the southern edge of the MMB) propagates from an initial spike north and updip along a curving ridge for about 2 km. The initial spike continued to grow in the November 14, 1993, event (located north of the October 20, 1992, event just beneath the hypocenter of the 1966 Parkfield earthquake), which shows little directivity, although there is a smaller patch of slip updip and to the south. In contrast, rupture for the December 20, 1994, event (located just south of the October 20, 1992, event) propagated north and slightly updip, creating a rough plateau in slip a few kilometers wide on a side. Directivity for this event also is to the north. Directivity for all three events points in the approximate direction of the 1966 hypocenter. Small pulses, which comprise a coda, are found on the STFs for several seconds after the initial impulsive event. Several tests based on the assumption that the average of all STFs from UPSAR for each event is an estimate of the true slip at the source suggest that the codas in the STFs are S waves from a long-duration source rather than uncorrected site response. An initiation phase is found on the array average for the November 14, 1993, and December 20, 1994, events. These precursory phases are the result of a spike in slip at the hypocenter. A value of 2.4--4 mm is obtained for Dc, the slip-weakening distance, by interpreting the initial spike as a critical patch. The few aftershocks for the October 20, 1992, event are distributed to the north and updip of the mainshock, but the November 14, 1993, event had a strong burst of aftershock activity that propagated to the north of its hypocenter at roughly the same depth. Aftershocks of the December 20, 1994, event are mostly updip. The November 14, 1993, event had the simplest slip distribution, appeared to be the most impulsive, and had the most active aftershock sequence and the greatest depth. If the eventual Parkfield earthquake initiates near the 1966 hypocenter, then the directivity of the three events studied here will have pointed to it. However, it is certainly possible that both the initiation of characteristic Parkfield shocks and the directivity of smaller events are controlled by fault properties on a larger scale such as by fault bends or jogs.