The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill mandates that the agent ingest and copy verbatim text from external, untrusted sources such as Amazon reviews, Trustpilot, Reddit, and social media comments. This creates a surface for indirect prompt injection where instructions embedded in these reviews could influence the agent's behavior or be written into campaign files.
Ingestion points: references/persona-research.md explicitly instructs the agent to spend time mining competitor reviews and social media threads for exact customer language.
Capability inventory: The skill has the capability to execute local bash scripts and write multiple files to the workspace/ directory.
Sanitization: The instructions require the agent to "Copy exact phrases — do not paraphrase," which increases the risk of copying malicious instructions directly into the script pipeline.
Boundary markers: The skill does not provide specific delimiters or instructions to the agent to distinguish between user sentiment data and potential control instructions within the mined text.
[COMMAND_EXECUTION]: The skill workflow relies on the execution of local shell scripts to manage creator profiles and process video assets.
Evidence: SKILL.md contains an explicit call to bash scripts/create-creator.sh <workspace> <creator-name> and references the use of scripts/build-hook-demo.sh for video assembly.

scrollclaw-persona