The Agent Skills Directory

[PROMPT_INJECTION]: A static detector flagged an 'ignore instructions' pattern in references/templates/classification.md. This is a false positive, as the text is part of a sample dataset used to test if a classification AI can identify such inputs as spam; it is not an instruction for the agent.
[PROMPT_INJECTION]: The skill possesses an indirect prompt injection surface as it analyzes project code and prompts to generate test suites.
Ingestion points: SKILL.md instructions specify reading project source code, system prompts, and test data.
Boundary markers: Absent. No specific delimiters are recommended for processing untrusted content during generation.
Capability inventory: The skill allows for writing .eval.ts files and importing application code into them.
Sanitization: Absent. No explicit sanitization of ingested content is described.
[DATA_EXFILTRATION]: While the skill accesses local project files to facilitate evaluation writing, this activity is confined to the local development environment and serves the tool's intended purpose. No patterns of external data transmission were found.

writing-evals