The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes law texts from the e-Gov API, which is a source of untrusted data.
Ingestion points: Untrusted data enters the context via scripts/fetch-law.sh and scripts/search-keyword.sh (documented in SKILL.md).
Boundary markers: The skill explicitly uses JSON string encoding as a pipeline boundary to prevent instructions from being treated as commands. It also provides specific behavioral instructions to the AI to treat retrieved text as data only.
Capability inventory: The skill uses curl for API access and local shell utilities (grep, sed, awk) for parsing. It does not perform sensitive file writes or system-level configuration changes.
Sanitization: Law texts are kept in raw JSON format to maintain the data/instruction boundary. A specific --text mode exists for human reading but is not the default for AI processing to avoid flattening the data.
[COMMAND_EXECUTION]: The skill uses local shell scripts to execute curl commands.
Evidence: scripts/fetch-law.sh and other scripts construct URLs and execute curl. However, input validation is performed on parameters like asof (date format check) and law_id (alphanumeric check in lib/source-url.sh) to mitigate common injection risks.
[SAFE]: All external communications target the official Japanese government domain laws.e-gov.go.jp.
Evidence: SKILL.md and script files correctly point to official API endpoints. No unauthorized third-party domains or suspicious downloads were detected.

jp-law