The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process content from untrusted external sources, including customer reviews, competitor ads from the Meta Ad Library, and social media posts.
Ingestion points: Data is retrieved via tools such as brand_knowledge_search (for reviews and facts) and ads_template_read (for competitor ads), as detailed in references/reading-the-brain.md.
Boundary markers: Instructions in references/prompts.md attempt to constrain the model with phrases like "Render only the text given below", but these are insufficient to prevent adversarial instructions embedded in the ingested content from influencing the agent.
Capability inventory: The agent has the ability to write to the local file system (file_write) and trigger image generation (ads_render), which could be targets for manipulation via indirect injection.
Sanitization: The playbook lacks explicit sanitization or filtering of external content before it is interpolated into prompts.
[EXTERNAL_DOWNLOADS]: The scripts/template_sheet.py script facilitates the retrieval of images from arbitrary remote URLs.
Evidence: The script uses urllib.request.urlopen to fetch image data from URLs provided as command-line arguments, which is a necessary part of the ad template preview functionality.
[COMMAND_EXECUTION]: The skill workflow involves executing several local Python scripts to perform image processing and layout tasks.
Evidence: The playbook and scripts/README.md document the execution of place_logo.py, template_sheet.py, and repair_patch.py using the python3 interpreter.

make-static-ads