The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user-supplied content such as ideas, decisions, and plans, which represents an ingestion surface for potential indirect prompt injection attacks.
Ingestion points: Data enters via the primary skill argument (SKILL.md).
Boundary markers: The skill uses --- delimiters to clearly separate agent analysis and question rounds from other context (references/questions.md).
Capability inventory: The agent is authorized to execute shell searches (rg, grep), write artifacts to the .tigerkit/ directory (artifact-paths.md), and call a local helper script (references/artifact-paths.md).
Sanitization: The instructions explicitly require quoting search globs and avoiding dangerous shell patterns (such as unquoted words starting with = in zsh) to mitigate command injection risks (SKILL.md).
[DYNAMIC_EXECUTION]: The skill utilizes dynamic execution for its visual question aid and artifact management.
Visual Aid: It generates an HTML file (questions.html) containing bundled JavaScript (html-theme.js). This script manually constructs ZIP files in-browser for user downloads using DataView and ArrayBuffer, designed specifically to avoid external network dependencies.
Local Scripting: The skill calls a local Python script, scripts/durable_artifacts.py, to check the status of durable artifacts (references/artifact-paths.md).
[COMMAND_EXECUTION]: The skill performs repository-local searches using rg (ripgrep) and grep. These commands are constrained by instructions to bind lookups to the verified repository root and use specific quoting practices to ensure safety against argument injection.

tk-grill