The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes shell commands using placeholders for user input, specifically grep -rni "<keywords>" "$MEMORY_DIR/". This pattern is susceptible to command injection if the input contains shell metacharacters such as semicolons, pipes, or backticks. It also uses dynamic shell execution with pwd and sed to construct paths to internal storage.
[DATA_EXFILTRATION]: The skill is designed to read from the $HOME/.claude/projects/ directory, which contains project-specific history and sensitive context. The workflow facilitates the extraction and potential external publishing of this content via the clawhub publish command, creating a risk of accidental data leakage despite manual quality gate instructions.
[PROMPT_INJECTION]: The skill ingests untrusted data from both user descriptions and internal project memory to generate new agent instructions. It lacks programmatic sanitization or boundary markers to mitigate indirect prompt injection during the extraction process. Ingestion points: User-provided pattern descriptions and files within the agent's project memory directory. Boundary markers: None identified in the workflow or shell command templates. Capability inventory: Shell command execution, file system writes, and the ability to spawn subordinate agents. Sanitization: The skill relies on non-enforced manual instructions to omit project-specific credentials or values rather than automated validation.

extract