The Agent Skills Directory

[PROMPT_INJECTION]: The skill provides instructions to bypass platform security filters and safety guidelines by using the --dangerously-skip-permissions flag with the opencode and claude CLI tools. This flag is explicitly documented to 'auto-approve file writes' and grant 'full autonomy,' effectively overriding the standard security model that requires human-in-the-loop confirmation for filesystem modifications. Additionally, the skill establishes a surface for indirect prompt injection by interpolating untrusted user input into shell commands.
Ingestion points: User-provided design prompts enter the agent context in SKILL.md and references/prompt-templates.md.
Boundary markers: Examples wrap prompts in double quotes, but no instructions are provided to sanitize or escape input to prevent command injection or shell escaping.
Capability inventory: The delegated tools (opencode, claude) have the capability to read and modify any file in the project.
Sanitization: The skill contains no instructions for validating, filtering, or escaping external content before execution.
[COMMAND_EXECUTION]: The skill's primary function is implemented through the execution of external shell commands. The lack of input sanitization combined with the use of safety-bypass flags significantly increases the risk of unauthorized command execution.
[REMOTE_CODE_EXECUTION]: The skill facilitates the execution of remote code by delegating tasks to external models and granting them the ability to write to the local filesystem autonomously. This enables remote-generated code to be integrated into the local development environment without manual oversight or approval.

gpt-design-bypass