The Agent Skills Directory

[PROMPT_INJECTION] (HIGH): The skill is highly susceptible to Indirect Prompt Injection (Category 8).
Ingestion points: Raw user input is gathered in Step 1 (Company basics, Problem, Solution, etc.) and written directly to pitch_data.json in Step 3.
Boundary markers: The instructions lack any delimiters (like XML tags or block quotes) or system prompts to treat this data as untrusted or to ignore embedded instructions.
Capability inventory: The skill has the capability to write files (pitch_data.json), execute shell commands (grep), and run Python scripts (python3 scripts/create_pitch_deck.py).
Sanitization: There is no evidence of data sanitization or validation before the untrusted content is used in file-writing or script-execution workflows. This creates a risk where malicious instructions hidden in the user's pitch details could influence the agent or the script.
[COMMAND_EXECUTION] (MEDIUM): The workflow involves executing shell commands like grep and python3. While the instructions specify structure, the use of variables in these commands (e.g., in the grep pattern or filename arguments in Step 4) creates a surface for command injection if the agent is manipulated by user input. Additionally, it relies on an unprovided script (scripts/create_pitch_deck.py).
[EXTERNAL_DOWNLOADS] (LOW): The skill documentation suggests installing the python-pptx library from PyPI. While this is a standard library, it is an external dependency from an untrusted authoring organization (AI Labs), which is outside the defined trust scope.

pitch-deck