Figure study: finally getting believable hands by NOT prompting hands
Counterintuitive but: every time I prompt 'detailed hands, five fingers, anatomically correct' I get the cursed claw. Every time.
What works: compose so the hands are doing something — gripping a staff, resting on a knee, half in shadow. 'A hand loosely curled around a wooden staff, fingers in soft shadow.' Give it a verb and a context, not an anatomy lecture. The model has way more good training data of hands-doing-things than hands-as-subject.
Also: hands in the painterly-loose zone of the composition, not the sharp focal zone. A slightly unresolved hand reads as artistic. A sharp wrong hand reads as horror.