The Agent Skills Directory

[PROMPT_INJECTION]: The skill defines patterns for sub-workflows that ingest untrusted data through Execute Workflow Trigger nodes, specifically highlighting integration with AI agents using fromAi() parameters. This architecture creates a surface for indirect prompt injection where malicious input could influence downstream agent behavior or workflow logic if not properly handled.
Ingestion points: Untrusted data enters the agent context via Execute Workflow Trigger nodes as described in SKILL.md (Sub-workflow inputs and outputs section).
Boundary markers: The documentation recommends using "Define Below" with explicit typed fields and suggests using sticky notes to document when no inputs are expected, which helps clarify the contract but does not inherently block malicious content.
Capability inventory: The skill describes workflows with the capability to perform HTTP requests, database operations (reads and writes), and dispatching notifications (Slack/SMTP) in SKILL.md (Stateful section).
Sanitization: The skill promotes best practices including returning consistent JSON shapes for errors, using explicit Set nodes to define return contracts, and normalizing data from storage layers (Implementation detail vs Interface) in SKILL.md and references/SUBWORKFLOW_PATTERNS.md.

n8n-subworkflows-official