Context. Differentiating between a true exoplanet signal and residual speckle noise is a key challenge in high-contrast imaging (HCI). Speckles result from a combination of fast, slow, and static wavefront aberrations introduced by atmospheric turbulence and instrument optics. While wavefront control techniques developed over the last decade have shown promise in minimizing fast atmospheric residuals, slow and static aberrations such as non-common path aberrations (NCPAs) remain a key limiting factor for exoplanet detection. NCPAs are not seen by the wavefront sensor (WFS) of the adaptive optics (AO) loop, hence the difficulty in correcting them.Aims. We propose to improve the identification and rejection of slow and static speckles in AO-corrected images. The algorithm known as the Direct Reinforcement Wavefront Heuristic Optimisation (DrWHO) performs a frequent compensation operation on static and quasi-static aberrations (including NCPAs) to boost image contrast. It is applicable to general-purpose AO systems as well as HCI systems.Methods. By changing the WFS reference at every iteration of the algorithm (a few tens of seconds), DrWHO changes the AO system point of convergence to lead it towards a compensation mechanism for the static and slow aberrations. References are calculated using an iterative lucky-imaging approach, where each iteration updates the WFS reference, ultimately favoring high-quality focal plane images.Results. We validated this concept through both numerical simulations and on-sky testing on the SCExAO instrument at the 8.2-m Subaru telescope. Simulations show a rapid convergence towards the correction of 82% of the NCPAs. On-sky tests were performed over a 10 min run in the visible (750 nm). We introduced a flux concentration (FC) metric to quantify the point spread function (PSF) quality and measure a 15.7% improvement compared to the pre-DrWHO image.Conclusions. The DrWHO algorithm is a robust focal-plane wavefront sensing calibration method that has been successfully demonstrated on-sky. It does not rely on a model and does not require wavefront sensor calibration or linearity. It is compatible with different wavefront control methods, and can be further optimized for speed and efficiency. The algorithm is ready to be incorporated in scientific observations, enabling better PSF quality and stability during observations.