Accurate sky description is essential for daylight research and practice, yet existing methods lie at two extremes. The CIE 15-type Standard General Sky requires full-dome luminance scans, which are rarely obtained, whereas most simulation engines offer only three or four sky presets that, while convenient, overlook important variation. We propose a camera-based middle ground: a 16-dimensional Sky Image Descriptor (SID) that preserves the intuitive CIE families (clear, intermediate and overcast) while retaining much gradational nuance. SID was trained on 21 490 public Sky Finder frames and evaluated using 44 calibrated HDR-LDR window-view sequences. HDR photographs supply absolute luminance and colour, and time-synchronized LDR videos capture cloud motion. In SID space, scenes form a smooth continuum from clear to overcast, and unsupervised clusters recover field-assigned CIE subtypes, thus combining scanner-level detail with three-class simplicity. Four continuous scene metrics (cloud coverage, optical flow, luminance and correlated colour temperature) are produced with each embedding, letting users quantify cloud fraction, motion and photometric balance without fitting the five CIE coefficients. The result is a practical taxonomy paired with detailed, parameterized descriptors that serve designers, engineers and researchers alike. All code, pretrained weights and annotated imagery are released as open-source, connecting camera-based workflows to scanner-based daylight research.