The Agent Skills Directory

[PROMPT_INJECTION]: The skill uses authoritative language to override standard agent behaviors, including directives that the skill MUST be active at all times and re-read after interruptions. It specifically includes a pattern to suppress self-defensive or critical statements ("If a sentence in a response serves to defend yourself... do not say it"), which is a known technique for bypassing AI safety guidelines.
[METADATA_POISONING]: The frontmatter and description fields use deceptive language claiming the skill is a "mandatory always-on" constraint that "MUST be loaded before every task," attempting to trick the agent's context management into prioritizing these instructions.
[INDIRECT_PROMPT_INJECTION]: The skill creates a surface vulnerable to indirect injection. 1. Ingestion points: The agent reads user requests, project source code, and repository instructions. 2. Boundary markers: None are defined to separate instructions from the data being processed. 3. Capability inventory: The agent is permitted to write files and execute shell commands for testing and linting. 4. Sanitization: No sanitization or escaping mechanisms are described for handling content from local files.
[COMMAND_EXECUTION]: The skill explicitly permits the agent to execute a variety of shell commands, including builds, tests, and validation scripts, which provides a functional attack surface if combined with other injection vectors.

sove-skills