The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process and synthesize data from external sources (such as user feedback, metrics, and competitive analysis) into product specifications, creating a potential surface for indirect prompt injection where malicious inputs could influence the agent's output.
Ingestion points: Untrusted data is ingested through the Pulse (usage metrics), Voice (user feedback), and Compete (competitive gaps) modules, as specified in the collaboration patterns defined in SKILL.md.
Boundary markers: The instructions do not prescribe explicit delimiters (e.g., [BEGIN DATA]) or specific warnings to ignore instructions embedded within these external inputs.
Capability inventory: The agent has the ability to write documents to the filesystem and execute shell commands for subagent orchestration as described in the multi recipe.
Sanitization: The framework relies on human-led judgment and review stages (mentioned in reference/modern-product-discovery.md) but lacks automated data sanitization or escaping mechanisms.
[COMMAND_EXECUTION]: The skill's 'multi' recipe involves the use of external CLI tools and instructs the agent to perform environment discovery using shell commands.
Evidence: reference/tri-engine-proposal.md contains a shell script for the agent to execute to detect the presence of tools like codex, agy, and claude by probing paths such as /usr/local/bin/ and running version checks.

spark