The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted datasets (JSONL, CSV, JSON) provided by the user, which presents a surface for indirect prompt injection attacks.
Ingestion points: The skill reads external data files in Phase 1a of SKILL.md and processes them to build failure mode taxonomies.
Boundary markers: There are no explicit instructions or delimiters defined to prevent the agent from obeying instructions that might be embedded within the dataset records.
Capability inventory: The skill possesses significant capabilities including local file system access (read/write to the error_discovery_data directory), running a local HTTP server, and spawning subagents to scan the corpus.
Sanitization: The instructions do not include mechanisms to sanitize or filter external content before it is interpreted by the LLM during analysis or subagent execution.
[COMMAND_EXECUTION]: The skill instructs the agent to execute a Python HTTP server using the standard library and to spawn background subagents for corpus-wide scanning. While these are functional requirements of the skill, they represent active process management within the environment.
[EXTERNAL_DOWNLOADS]: The skill recommends using well-known Content Delivery Networks (CDNs) to fetch standard JavaScript libraries, such as marked.js for markdown rendering and highlight.js or Prism for syntax highlighting. These are recognized services for frontend web development.

error-discovery