The Agent Skills Directory

[COMMAND_EXECUTION]: The evaluation phase (Phase 7) described in scripts/generate.sh and references/eval/phase.md instructs the agent to run all bash scripts matching tests/test-*.sh and execute shell commands from the assert_sh array in the target skill's test files. This enables execution of arbitrary shell code defined in external data files.
[PROMPT_INJECTION]: The skill processes untrusted data (target skill files) and provides them to subagents during evaluation. 1. Ingestion points: SKILL.md and reference files in the target skill directory. 2. Boundary markers: Absent in subagent prompts for processing target skill content. 3. Capability inventory: Bash, Read, Write, Edit, Grep, Glob, Skill tools. 4. Sanitization: Absent; content is provided directly to subagents.

skills-engineering