The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/gate-check.mjs implements an arbitrary command execution capability. It parses GATES.md files for a CHECK: attribute and passes the resulting string directly to spawnSync with shell: true. This allows the AI agent (or an attacker via the agent) to execute any command on the host system without validation.
[PERSISTENCE]: The skill includes scripts/install-hooks.mjs, which modifies the AI agent's environment configuration. It writes a persistent "Stop" hook into .claude/settings.local.json or the global ~/.claude/settings.json. This hook executes scripts/stop-hook.mjs to evaluate whether the agent is permitted to end its turn, effectively overriding the platform's default session termination behavior.
[PROMPT_INJECTION]: The SKILL.md file contains instructions designed to override the agent's internal resource management and safety heuristics. Specifically, it commands the agent to "Ignore resource anxiety," "enforce completion through gate files," and "not stop until every gate is checked." These instructions are intended to prevent the agent from terminating a session prematurely, which directly conflicts with typical platform-level constraints.
[INDIRECT_PROMPT_INJECTION]: The skill creates a significant attack surface for indirect prompt injection. Because the agent is instructed to create CHECK: shell commands based on task descriptions, a malicious user could provide a task containing embedded instructions that result in the creation and execution of dangerous shell commands.
Ingestion points: User-provided task descriptions are processed by the agent to generate GATES.md files (SKILL.md).
Boundary markers: None. There are no instructions to sanitize or escape user-controlled data before it is written into CHECK: shell commands.
Capability inventory: Arbitrary shell execution via spawnSync in scripts/gate-check.mjs and persistent agent behavior modification via scripts/install-hooks.mjs.
Sanitization: Absent. The scripts perform no validation or filtering on the content extracted from gate files before execution.
[EXTERNAL_DOWNLOADS]: The documentation in README.md and SKILL.md directs users and agents to install the skill by cloning a repository from GitHub (https://github.com/Leonxlnx/unlazy) or using npx skills add. While these target well-known platforms, the skill facilitates the introduction of local executable scripts into the agent's environment.

unlazy