The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied parameters through $ARGUMENTS and reads draft evaluation patterns from the local filesystem. These inputs are used to cluster themes and generate adversarial examples which are then passed to the /evals-implement skill for automated code generation of graders. The absence of explicit input sanitization or boundary markers when handling these external data sources creates a surface for indirect prompt injection that could affect the integrity of the resulting evaluation datasets.
Ingestion points: User input captured via $ARGUMENTS in SKILL.md; draft evaluation criteria patterns read from .adlc/memory/evals/EVAL-*.md.
Boundary markers: No delimiters or instructions to ignore embedded commands are present when processing the external criteria or user arguments.
Capability inventory: The skill performs local file operations (reads and writes) within the project directory and triggers external skill execution (/evals-implement).
Sanitization: No evidence of escaping or validation is present for the criteria being processed or the arguments being passed.
[DYNAMIC_EXECUTION]: The bash setup script (scripts/bash/setup-evals-clarify.sh) utilizes a Python heredoc to format environment variables into a JSON object. This is a standard utility pattern for ensuring consistent data output across different shell environments and does not incorporate untrusted external input into the execution string.

evals-clarify