The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes transcription data from the NDL Kokkai API, which contains text from various third parties (MPs, witnesses, etc.). This represents a potential surface for indirect prompt injection.
Ingestion points: The speech and meetingRecord fields retrieved by the scripts in the scripts/ directory and processed as described in SKILL.md.
Boundary markers: The skill explicitly instructs the agent to treat retrieved text as raw data and maintain JSON boundaries, rather than flattening text into a single prompt. It also warns against following commands embedded within the speech text.
Capability inventory: The skill provides bash wrapper scripts (fetch-meeting.sh, list-meetings.sh, etc.) that use curl and jq. It does not have write access to sensitive files or generic code execution capabilities.
Sanitization: The skill relies on behavioral guidance in SKILL.md, instructing the AI to treat external content as untrusted data and ignore any instructions contained within it.
[COMMAND_EXECUTION]: The skill uses local bash scripts to interact with the external API. This is the primary function of the skill and is implemented using standard tools like curl and jq.
Evidence: The scripts/ directory contains various .sh files that wrap API calls to https://kokkai.ndl.go.jp/api/.
[SAFE]: The skill communicates exclusively with the official Japanese National Diet Library (NDL) API at kokkai.ndl.go.jp, which is a well-known and trusted government service for parliamentary records. No evidence of obfuscation, persistence, or malicious behavior was found.

jp-diet-minutes