The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted text from meeting transcripts and interpolates it directly into LLM prompts without sanitization or robust boundary markers.
Ingestion points: summarize.py reads user-provided file content or standard input via the read_input function.
Boundary markers: The prompt construction in summarize_openai and summarize_gemini uses simple newline separators instead of secure delimiters or specific instructions for the LLM to ignore embedded commands within the content.
Capability inventory: The script has the capability to write processed content to the local filesystem (save_output) and execute external binaries (subprocess.run).
Sanitization: No sanitization or content filtering is performed on the input data before it is interpolated into the system and user messages.
[COMMAND_EXECUTION]: The script executes external CLI tools to handle PDF text extraction and interface with the Gemini model.
Evidence: summarize.py invokes pdftotext and a gemini binary using subprocess.run with argument lists.
Context: While the use of argument lists mitigates shell injection, these tools are not declared in the SKILL.md metadata requires section, which may lead to runtime errors or interactions with unexpected binaries in the user's environment.
[METADATA_POISONING]: The skill documentation references non-existent or futuristic model versions, which may be deceptive regarding the actual capabilities of the tool.
Evidence: SKILL.md and summarize.py mention gpt-5.4 and gemini-3.1-pro-preview as supported models, neither of which are currently available versions of these services.

jiyao-summary