The Agent Skills Directory

[PROMPT_INJECTION]: The skill uses directive language such as "IMMEDIATELY invoke the script" and "Do NOT analyze first" in both SKILL.md and CLAUDE.md. These instructions are designed to override the agent's default behavior of analyzing tasks before execution, which can be used to bypass safety reflections.
[COMMAND_EXECUTION]: The skill requires the agent to run a shell command (python3 -m skills.decision_critic.decision_critic) in SKILL.md. The code for this script is missing from the provided files, making its behavior unverifiable and potentially dangerous if the local environment is compromised or if the script contains malicious logic.
[INDIRECT_PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection because it accepts untrusted input.
Ingestion points: The <decision text> placeholder in SKILL.md is intended to be filled with user-provided data.
Boundary markers: The input is wrapped in single quotes within the command string, which is an insufficient boundary that can be escaped in many shell environments.
Capability inventory: The skill has the capability to execute shell commands via the <invoke> tag.
Sanitization: The instructions do not define any sanitization, escaping, or validation logic for the user-supplied decision text before it is passed to the command line.

decision-critic