The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests external, untrusted data including meeting transcripts and client intake forms. It includes a specific safety instruction in references/meeting-notes.md to prevent the execution of instructions found within these records, ensuring that the agent treats speech strictly as text to summarize rather than actions to perform.\n- [DYNAMIC_EXECUTION]: The documentation for Obsidian vault management (references/obsidian.md) includes the obsidian eval command, which allows JavaScript evaluation within the Obsidian app. This is presented as a developer tool for plugin and theme verification. No instructions exist that would cause the agent to execute untrusted code through this interface.\n- [DATA_EXPOSURE]: The skill incorporates defensive measures against accidental data exposure. The scripts/multi-client.ts utility scans for hardcoded secrets and verifies that sensitive files like .env are excluded from version control. Additionally, the references/vault.md file defines a protocol for masking client credentials in logs and transcripts using a tenant-scoped redaction format.\n- [COMMAND_EXECUTION]: The skill utilizes local scripts (scripts/multi-client.ts and scripts/sku-packager.ts) to automate agency operations such as workspace isolation audits and pricing calculations. These scripts perform safe, local file and data operations without external network dependencies.

ops