The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes various shell commands to render output and perform visual checks, including bash, tinymist, pdftotext, screencapture, and npx. Notably, it calls shell scripts located in parent and sibling directories (e.g., ../craft/scripts/goal-self-send.sh and ../../skills/look-at/scripts/look_at.sh), assuming a specific directory structure for its dependencies.
[DYNAMIC_EXECUTION]: The skill instructions provide a mechanism to spawn "general-purpose" subagents using the Agent() function to perform visual scoring tasks when the primary vision tool is unavailable. Additionally, it dynamically generates temporary Typst source files (test-slide.typ) from templates to perform isolated rendering tests.
[INDIRECT_PROMPT_INJECTION]: The skill constructs prompts for external vision models (Gemini, GPT-5.4, and Claude) by interpolating content from task specifications like SPEC.md or PLAN.md. This creates a surface for indirect prompt injection, where malicious instructions inside those documents could potentially influence the reviewer model's output.
Ingestion points: Content from task specifications, project plans, and user requests is used to populate placeholders in vision goal templates.
Boundary markers: The skill uses markdown headers and specific structured formats (e.g., "## Check These Specifically") within its templates to separate context from instructions.
Capability inventory: The skill possesses capabilities for file system writes (cat), shell execution (tinymist, pdftotext), and subagent spawning.
Sanitization: There is no explicit evidence of sanitization or escaping applied to the interpolated task data before it is sent to the vision models.

visual-verify