The Agent Skills Directory

[SAFE]: The skill implements secure secret management by using SOPS for local encryption of environment files and enforcing restricted file permissions (chmod 600) on sensitive data paths like ~/.env.skill-studio.
[COMMAND_EXECUTION]: Subprocess calls are used for legitimate local operations including secret encryption/decryption via sops, running the project's test suite via pytest, and executing environment preflight checks. All commands target established tools or the project's own modules.
[EXTERNAL_DOWNLOADS]: Network operations target well-known and reputable services for their intended functionality, such as OpenRouter and Anthropic for LLM access, Google Gemini for model verification, and Daily.co for provisioning voice interview rooms.
[INDIRECT_PROMPT_INJECTION]: The skill can ingest Claude Code session transcripts to seed new design proposals. This ingestion point is protected by a deterministic extraction process in src/skill_studio/ingest/transcript.py that uses regular expressions and JSON parsing to compress data before it is processed by an LLM, preventing the direct execution of arbitrary instructions found in user data.
[SAFE]: The skill architecture follows least-privilege principles, keeping session data within user-controlled directories (~/.skill-studio) and avoiding hard-coded credentials or absolute system paths.

skill-studio