The Agent Skills Directory

[SAFE]: The skill defines a collaborative methodology that ensures the user remains in full control. It contains explicit instructions that forbid the agent from performing any autonomous investigation, stating: 'Never execute an investigation step without the user's explicit go-ahead.'\n- [COMMAND_EXECUTION]: While the skill involves the use of system tools like Bash and Grep for data exploration, the risk is minimized by mandatory confirmation steps where the agent must present its proposed plan and wait for user approval before execution.\n- [DATA_EXPOSURE_AND_EXFILTRATION]: Research reports and audit logs are stored locally in the ~/.ai-software-agency/ directory. These are standard file operations for project documentation and do not involve unauthorized exfiltration or access to sensitive system credentials.\n- [INDIRECT_PROMPT_INJECTION]: The skill is intended to process external information such as code and research data. Potential indirect injections are mitigated by the 'sparring partner' stance, which requires the agent to interpret data alongside the user and avoid silent shifts in hypothesis.

research-sparring