The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data during its discovery and intake phases, creating a potential vector for indirect instructions to influence agent behavior or poison project artifacts.
Ingestion points: The skill ingests visitor research, exit surveys, support tickets, and chat logs as described in the 'Intake' section and Phase 1 (Discovery).
Capability inventory: The skill performs repeated file writes to multiple project artifacts in the local docs/ directory, including WEBSITE.md, OFFER.md, MARKETING.md, EXPERIMENTS.md, POSITIONING.md, METRICS.md, and GROW-WEBSITE-PLAN.md.
Sanitization: There are no specific instructions or mechanisms for sanitizing, validating, or escaping the external research data before it is interpolated into project documentation.
Boundary markers: The instructions lack boundary markers or warnings to the agent to disregard potentially malicious instructions embedded within the ingested visitor data.
[EXTERNAL_DOWNLOADS]: The skill provides instructions to suggest the installation of external tools using a package runner.
Evidence: The 'Operating Rules' specify that if a dependent skill is missing, the agent should offer to run npx skills add wondelai/skills/<slug> --global.
Context: The suggested packages are resources belonging to the skill's author ('wondelai') and are documented here as standard vendor functionality.

grow-website