The Agent Skills Directory

[PROMPT_INJECTION]: The skill presents an indirect prompt injection surface by design, as it ingests and processes untrusted external content like URLs, documents, and transcripts to generate presentation materials.
Ingestion points: Step 1 uses the WebFetch tool to read user-provided URLs and imports uploaded files via the import_attachment tool in SKILL.md.
Boundary markers: The workflow does not include specific instructions or delimiters to isolate the agent from potentially malicious instructions embedded in the source data.
Capability inventory: The skill can write files to the local directory through run_python and dispatch sub-agents across all scripts.
Sanitization: Although the run_python tool includes a linting and sanitization step for output, there is no evidence of NLP-level sanitization for the input data derived from external sources.
[COMMAND_EXECUTION]: The skill utilizes the run_python tool to execute dynamically generated Python code snippets for managing files and data within the presentation workspace.
Evidence: The instructions in SKILL.md explicitly direct the agent to generate and execute code such as write_text("specs/brief.md", content) and write_json for updating deck.json.
Context: These operations are performed within a sandboxed environment intended to restrict unauthorized imports and network access, minimizing the risk of arbitrary code execution outside the workspace.

sdpm-vibe