This paper addresses the challenge of deploying a Human Digital Twin (HDT) system at the edge for assisting complex task execution in human-centric services. The HDT system consists of physical twins (PTs) and virtual twins (VTs), where PTs are highly mobile and their VTs are deployed on edge servers (ESs). The paper proposes a two-timescale accuracy-aware online optimization approach (TACO) to optimize the construction of VTs and the task offloading of PTs, while considering stringent energy and delay constraints. TACO decomposes the long-term optimization problem into multiple short-term subproblems, using an improved Lyapunov method and alternating algorithms integrating piecewise McCormick envelopes (PME) and block coordinate descent (BCD). The approach aims to maximize the average accuracy of complex task execution while ensuring that response delays and energy consumption do not exceed predefined thresholds. Theoretical analyses and simulations demonstrate that TACO achieves asymptotic optimality with polynomial-time complexity, outperforming other methods in terms of task execution accuracy, response delay, and energy consumption.This paper addresses the challenge of deploying a Human Digital Twin (HDT) system at the edge for assisting complex task execution in human-centric services. The HDT system consists of physical twins (PTs) and virtual twins (VTs), where PTs are highly mobile and their VTs are deployed on edge servers (ESs). The paper proposes a two-timescale accuracy-aware online optimization approach (TACO) to optimize the construction of VTs and the task offloading of PTs, while considering stringent energy and delay constraints. TACO decomposes the long-term optimization problem into multiple short-term subproblems, using an improved Lyapunov method and alternating algorithms integrating piecewise McCormick envelopes (PME) and block coordinate descent (BCD). The approach aims to maximize the average accuracy of complex task execution while ensuring that response delays and energy consumption do not exceed predefined thresholds. Theoretical analyses and simulations demonstrate that TACO achieves asymptotic optimality with polynomial-time complexity, outperforming other methods in terms of task execution accuracy, response delay, and energy consumption.