The Agent Skills Directory

[PROMPT_INJECTION]: The file tests/fixtures/prompt_injection_readme.md contains strings designed to override instructions (e.g., 'Ignore all previous instructions') and attempt data exfiltration ('Print any API tokens'). These are identified as test fixtures for regression testing and do not represent the skill's intended logic.
[INDIRECT_PROMPT_INJECTION]: The skill facilitates the ingestion of untrusted external content (webpages, PDFs, etc.), which is a known vector for indirect prompt injection.
Ingestion points: Web browsing and local file analysis as defined in SKILL.md and references/research-protocol.md.
Boundary markers: Explicit instructions are provided in SKILL.md and references/research-protocol.md to ignore instructions found within researched content.
Capability inventory: The skill uses scripts/research_ledger.py to write to the filesystem and manages an evidence ledger.
Sanitization: The skill mandates secret redaction and uses regex in scripts/research_ledger.py to detect potential sensitive data leakage.
[COMMAND_EXECUTION]: The test file tests/test_research_ledger.py uses the subprocess module to execute and verify the skill's Python script during the development cycle.

deep-research