The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill's curation and annotation workflows, described in rules/curation-add-workflow.md and rules/curation-annotation.md, involve ingesting external content from URLs and files for processing by sub-agents.
Ingestion points: The WebFetch tool is used to pull content from external articles, documentation, and research papers for quality scoring and test query generation.
Boundary markers: The instructions do not specify the use of strict delimiters or explicit instructions for the agent to ignore instructions embedded within the ingested content.
Capability inventory: The skill possesses the ability to read and write files, perform network searches, and fetch web content, providing a significant capability surface if a prompt injection is successful.
Sanitization: No explicit sanitization or content filtering is implemented before the untrusted data is processed by the agent's evaluation logic.
[DYNAMIC_CONTEXT_INJECTION]: The file scripts/backup-golden-dataset.md utilizes the !command`` syntax to execute shell commands when the skill content is loaded by a compatible agent platform.
Evidence: The file contains triggers for date, find . -type d -name "*golden*" ..., python --version, and pwd.
Purpose: These commands are used to auto-detect the environment (current directory and Python version) and identify potential dataset locations to populate documentation templates for the user.
Assessment: These specific commands are benign and serve legitimate project-specific tooling purposes, though they execute automatically at skill load time.
[EXTERNAL_DOWNLOADS]: The skill references several external resources and dependencies for validation and storage functionality.
Evidence: SKILL.md identifies json-schema.org, zod.dev, langfuse.com, github.com/pgvector/pgvector, and postgresql.org as authoritative sources for schema validation, duplicate detection, and storage mechanics.
Assessment: These references target well-known and trusted services/organizations and are documented neutrally for background context and upstream coverage.

golden-dataset