The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted external data from the web (via BrowserOS) and public review platforms such as G2 and Capterra, which are attacker-controllable sources.
Ingestion points: Web research and 'Review mining' tasks described in SKILL.md where the agent is instructed to gather competitor strengths, weaknesses, and common complaints.
Boundary markers: Although the instructions explicitly state that web content should be 'treated as hostile', there are no technical delimiters (like XML tags or clear separators) instructed for the ingested data to prevent the agent from obeying embedded instructions.
Capability inventory: The skill framework provides the agent with terminal access via the ExecutionProxy Interface and persistent file-writing capabilities through the MemoryStore Interface.
Sanitization: The protocol includes a requirement to redact PII and secrets, but it does not specify methods for sanitizing malicious natural language instructions that might be contained within the scraped reviews.
[COMMAND_EXECUTION]: The 'GLOBAL PROTOCOLS' section in SKILL.md mandates the use of an ExecutionProxy Interface (prefixing commands with rtk) for terminal actions, specifically mentioning execution of tests such as rtk npm test.
[NO_CODE]: The skill package is composed entirely of Markdown instruction files (SKILL.md, references/content-architecture.md, references/templates.md) and does not bundle any executable scripts or binaries.

competitor-alternatives