The Agent Skills Directory

[SAFE]: The skill implements a 'Retrieved Evidence Boundary' that explicitly instructs the agent to treat external content (web pages, files, logs) as data rather than instructions, preventing indirect prompt injection attacks.
[SAFE]: Sensitive user inputs are handled with security best practices, including the use of isolated directories with restrictive permissions (0700 for directories and 0600 for files).
[SAFE]: The skill maintains research isolation by recommending dedicated workspaces or linked worktrees for experimental mutations, ensuring that unrelated project work is not modified.
[SAFE]: External research sources and repositories are documented for provenance and design reference in accordance with the trust scope rule, with no direct execution of remote code detected.
[SAFE]: File system operations for state persistence and reporting are governed by strict ownership and path verification checks, including checking for symlinks and git-ignore status, to prevent unauthorized or accidental file writes.

tk-research