The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill acts as a guided Q&A interface that ingests user requirements to generate agent configurations and system prompts. Because these agents are designed to process user input and interact with the file system or web, they represent a surface for indirect prompt injection.
Ingestion points: User descriptions of agent tasks and personas are processed to generate SKILL.md frontmatter and system prompts, as described in the 'Creation Workflow' in SKILL.md.
Boundary markers: The skill recommends structured system prompts (using XML tags as suggested in references/agent-patterns.md) which can serve as delimiters, though it does not mandate specific sanitization logic for the agents it helps create.
Capability inventory: The skill documents configuration for powerful tools such as bash, edit, write, task (subagent spawning), and webfetch in references/opencode-config.md.
Sanitization: The skill actively promotes security by instructing users to avoid bash: "*" : "allow" and recommends a default-deny policy for skills combined with explicit allowlists, significantly hardening the resulting agents against exploitation.

opencode-design-agents