The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes bash commands to initialize complex session directory structures and manages its own dependencies by invoking pip install for libraries like weasyprint and python-pptx during the session.
[EXTERNAL_DOWNLOADS]: The skill uses WebFetch to download content from user-specified URLs to be used as primary source material for the brainstorming agents.
[REMOTE_CODE_EXECUTION]: Production agents (Archivist and Presentation Agent) generate Python scripts (build_capsule.py, build_presentation.py) based on the session's 'Vision Document.' Since this document is synthesized from potentially untrusted external inputs (files and URLs), it creates a risk where malicious patterns in the source data could be translated into executable code within the build scripts.
[PROMPT_INJECTION]: The skill exhibits vulnerability to indirect prompt injection due to its processing of external data.
Ingestion points: SKILL.md (Action 1: Capture Sources) and the PRD action ingest data from local files, external URLs, and user descriptions.
Boundary markers: The agent prompts do not utilize clear delimiters or instructions to ignore instructions embedded within the processed content.
Capability inventory: The orchestrator and agents have the ability to execute bash commands, perform network fetches, and generate/execute Python code.
Sanitization: There is no evidence of sanitization, filtering, or validation of the ingested content before it is processed by the generative agents or used in script construction.

ideation