The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The code snippet provided for Claude AI integration demonstrates a vulnerability surface for indirect prompt injection if the processed data is untrusted.
Ingestion points: The analyze_with_claude function in SKILL.md takes a content string and passes it directly into the messages array.
Boundary markers: The example lacks delimiters (e.g., XML tags or triple quotes) or specific instructions to the model to ignore instructions embedded within the content variable.
Capability inventory: The script demonstrates tool calling (provide_analysis) which could be manipulated by adversarial input.
Sanitization: There is no evidence of escaping, filtering, or validation of the input data before it is sent to the LLM.
[SAFE]: The skill provides standard documentation for project environment variables using recognizable placeholders (e.g., sk-ant-..., eyJ...) rather than actual secrets.
[SAFE]: All command-line examples for deployment and testing (gcloud run, pytest, npm) are routine development operations consistent with the skill's stated purpose.

project-guidelines-example