The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from external sources such as SRT subtitle files and CSV spreadsheets, as well as project-level metadata like composition and layer names. These inputs are reflected in the agent's context and could be used to influence its reasoning if they contain adversarial instructions.
Ingestion points: Subtitle text in srt-import.jsx, spreadsheet data in comp-from-csv.jsx, and project item names in project-overview.jsx.
Boundary markers: The skill does not implement delimiters or safety instructions when presenting this data to the agent.
Capability inventory: The agent has access to Bash (osascript), Write, and Read tools.
Sanitization: No significant sanitization or validation is performed on the ingested text before it is processed by the agent.
[DYNAMIC_EXECUTION]: The skill creates and executes scripts dynamically at runtime. It specifically instructs the agent to write custom JSX code to a file (ae-action.jsx) and then execute it via a shell command. While this is the intended design for After Effects automation, it represents a flexible execution model.
Evidence: SKILL.md Step 4 and Step 6 detail the process of writing and then running action scripts.
[COMMAND_EXECUTION]: The scripts/runner.sh script generates JXA (JavaScript for Automation) commands by directly interpolating shell variables into strings passed to osascript. This creates a potential injection vector if the input variables, such as file paths, were to contain characters that break the JXA string context.
Evidence: The construction of JXA_SCRIPT in runner.sh using "$AE_APP" and "$SCRIPT_PATH" without JXA-specific escaping.

after-effects