The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and review untrusted external data (project copy files, JSON dictionaries, and markdown docs). This creates an inherent surface for indirect prompt injection. However, the skill implements strict instructions for the agent to 'propose, never apply' changes in a table format, providing a human-in-the-loop checkpoint that mitigates the risk of the agent obeying instructions embedded in the analyzed copy.
Ingestion points: Reads project files including copy.json, package.json, .tsx components, and .md documentation.
Boundary markers: Instructions mandate the use of a 'proposal table' in reports, separating the analysis from file modifications.
Capability inventory: The skill uses standard file system tools (ls, grep) and a local Node.js linter script to perform its task.
Sanitization: The provided linter script (copy-lint.mjs) uses strict regular expressions to identify specific language patterns without interpreting the text as commands.
[DYNAMIC_EXECUTION]: The copy-lint.mjs script contains a --self-test mode that utilizes spawnSync to run itself against local test fixtures. This execution is limited to validating the linter's own logic during development or installation and does not involve executing external or untrusted code.
[SAFE]: The linter script relies solely on Node.js built-in modules (node:fs, node:path, node:child_process, node:url) and does not perform any network operations, access sensitive user credentials, or utilize obfuscation techniques.

visant-copy