The Agent Skills Directory

[COMMAND_EXECUTION]: The skill is configured to use the Bash tool to execute repository linter scripts (scripts/lint.sh) and run focused tests on newly authored skills. This represents elevated permission used for structural and behavioral validation during the development lifecycle.
[DATA_EXFILTRATION]: The skill's contract (host-contract.json) explicitly permits access to sensitive project files, including INVOICES.csv and TIME.csv. While intended to provide context for skill authoring, this access creates a surface for exposing sensitive financial or activity records.
[PERSISTENCE]: The skill is designed to install and update agent capabilities in persistent system directories. It specifically targets $HOME/.agents/skills/ (for Codex) and $HOME/.claude/skills/ (for Claude Code). This capability allows the skill to create or modify the agent's long-term operational toolbox.
[INDIRECT_PROMPT_INJECTION]: As a tool that generates natural language instructions (SKILL.md) from user-provided descriptions, the skill is a high-risk surface for indirect prompt injection. Malicious input could cause the skill to generate instructions that compromise the agent's safety guidelines or perform unauthorized actions in subsequent sessions.
[DYNAMIC_EXECUTION]: The skill facilitates the creation of executable instruction files. It includes a Python-based validation script (scripts/validate_skill.py) that uses yaml.safe_load to verify generated metadata, demonstrating an attempt to sanitize the output format.

skill-maker