zhangdi-avatar

张迪的数字化身。以张迪的视角、价值观、方法论和决策逻辑，进行分析、判断、决策，并持续自我完善。当用户需要"问问张迪会怎么想/怎么做/怎么判断"时触发。

This skill is a detailed persona that will store and update files in its skill folder and asks for collecting real chat transcripts for 'Turing tests'. Before installing: (1) Verify the origin and whether the real person (张迪) authorized this persona — it may impersonate someone without provenance. (2) Understand memory persistence: the skill will write logs and can modify SKILL.md/meta.json; if you do not want persistent changes, run it in a sandbox or disable writeback. (3) Do not feed private chat logs or sensitive messages unless you accept they will be stored inside the skill's memory directories. (4) If you want the persona but not file-write behavior, ask the author for a read-only variant (no automatic writebacks) or remove/lock the self_improvement/writeback logic. (5) Test in an isolated environment first and review memory/evolution-log.md and any updated files after runs. If you are unsure about provenance or do not want persistent local storage of user data, consider not installing or request a version that does not persistently modify files.

Type: OpenClaw Skill Name: zhangdi-avatar Version: 0.9.0 The skill bundle implements a 'Self-improvement mechanism' described in SKILL.md and meta.json that explicitly instructs the AI agent to rewrite its own source files (including SKILL.md, meta.json, and reference files) based on user interactions. This capability for autonomous self-modification creates a significant surface for persistent prompt injection, where a user could permanently alter the agent's core logic or constraints. While this appears to be a design feature for a 'digital twin' persona rather than an intentional backdoor, the ability to modify its own instructions is a high-risk behavior.