The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill instructs the agent to dynamically generate Python scripts (e.g., build.py, upgrade.py) and shell heredocs to handle data parsing and API interaction. These scripts are created at runtime in the /tmp directory and then executed to produce payloads for the Google Slides API.
[INDIRECT_PROMPT_INJECTION]: The workflow ingests untrusted Markdown data from a user-controlled slide.md file. This content is used to populate the generated Python scripts and shell commands, creating a risk that malicious instructions embedded in the Markdown could influence the agent's behavior.
Ingestion points: User-provided slide.md file processed in Step 3.
Boundary markers: None identified in the workflow to separate untrusted user data from agent-generated logic.
Capability inventory: Includes shell command execution (gws, python3, curl), file system writes (/tmp), and network access to Google Workspace APIs.
Sanitization: The skill lacks explicit instructions for sanitizing or escaping user-provided strings before they are interpolated into executable contexts.
[COMMAND_EXECUTION]: The agent executes multiple shell commands involving the gws CLI and python3 that incorporate variables derived from user input, such as the presentation title and slide identifiers, particularly in Steps 4 and 7.
[EXTERNAL_DOWNLOADS]: Step 7 uses curl to download slide thumbnails from URLs provided by the Google Slides API. While the destination (Google) is a well-known service, the interaction involves fetching remote content based on dynamically retrieved API data.

p-md-to-slides