The Agent Skills Directory

[PROMPT_INJECTION]: The skill processes untrusted image data via OCR and vision tools, creating a surface for indirect prompt injection where malicious instructions embedded in images could influence the agent's behavior.
Ingestion points: Image files are ingested and processed by multiple tools, including glance, detect, ground, and the long_screenshot_ocr.py script.
Boundary markers: The prompt logic in long_screenshot_ocr.py includes instructions to the vision model to avoid summarization or inference ("Do not summarize, rewrite, translate, or infer clipped text"), providing some protection against accidental instruction following.
Capability inventory: The skill environment supports subprocess execution (launching a browser via html_shot.py and the glance CLI via long_screenshot_ocr.py) and filesystem access for reading and writing image artifacts and transcripts.
Sanitization: Extracted OCR text is merged into markdown or JSON files without specific sanitization or escaping of the content found within the source images.
[COMMAND_EXECUTION]: The skill utilizes subprocesses to perform its core functions, such as capturing screenshots and calling vision utilities.
Evidence: The script scripts/html_shot.py uses subprocess.Popen and subprocess.run to launch and control a Chrome-family browser (Chrome, Chromium, or Edge) in headless mode for rendering HTML.
Evidence: The script scripts/long_screenshot_ocr.py uses subprocess.run to execute the glance vision CLI, resolving the executable path via the system PATH or relative repository locations.
Context: These executions are documented parts of the skill's intended functionality and are used to provide visual capabilities to the agent.

vision-tools