The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external files (target skills) which creates an attack surface for indirect prompt injection. However, the skill implements proactive defenses.
Ingestion points: The skill reads SKILL.md, rules, and reference files from target directories specified by the user as seen in SKILL.md and rules/repair-workflow.md.
Boundary markers: The rules/scenario-design.md file and assets/evals/skill-tester-cases.jsonl (e.g., ID: adversarial-retrieved-instruction) include explicit requirements to detect and reject adversarial instructions embedded in retrieved content.
Capability inventory: The skill utilizes edit and execute capabilities to perform repairs and run local validation scripts provided within the skill package, such as scripts/validate-skill.mjs.
Sanitization: The rules/repair-workflow.md enforces a 'baseline -> repair -> recheck' cycle and 'Safe deletion gate', ensuring that edits are minimal, verified against original safety guards, and do not impact files outside the target skill's proven ownership.

skill-tester