The Agent Skills Directory

[COMMAND_EXECUTION]: The skill requests and uses the Bash tool to perform local file system operations.
Its cli.js and install.sh scripts manage the installation and uninstallation of the skill files within the designated ~/.claude/skills/ directory.
The 'Nuwa/Creator' transformation dynamically generates new SKILL.md files based on user-requested personas, writing them to the local skills directory.
[EXTERNAL_DOWNLOADS]: The skill utilizes WebFetch and WebSearch tools as part of its research phase for public figure 'distillation'.
This process is governed by a 'Multi-Agent' architecture (references/distillation.md) that isolates different research dimensions and uses a triple-verification methodology to ensure the accuracy of gathered public information.
[PROMPT_INJECTION]: The skill possesses a surface for Indirect Prompt Injection because it is designed to ingest untrusted external data such as chat logs and social media profiles to create 'digital clones'.
Ingestion points: references/distillation.md and personas/fenshen.md (chat records, social media URLs via WebFetch).
Boundary markers: The framework includes mandatory 'Information Confirmation' and 'Result Confirmation' checkpoints (references/execution.md) where the agent must stop and wait for user approval before continuing.
Capability inventory: Access to Bash, Write, Edit, and Read tools.
Sanitization: The skill employs a 'Role Guard' and 'IS NOT' check system (references/quality.md) to detect and remove content that violates the persona's defined constraints or safety boundaries before outputting to the user.
[SAFE]: Extensive security documentation and '品味守则' (Taste Rules) are provided within the skill's reference files to prevent the fabrication of quotes or the misuse of gathered information. All external interactions are user-initiated and follow standard Claude Code skill implementation patterns.

wukong