The Agent Skills Directory

[PROMPT_INJECTION]: The references/bootstrap-monorepo.md file contains a 'Meta-Prompt' that explicitly assigns a persona ('Principal Software Architect') and instructs the agent to act autonomously while bypassing human interaction ('The human will not touch any code. You execute everything autonomously'). This pattern attempts to override standard safety guidelines regarding human-in-the-loop verification and agent autonomy.
[EXTERNAL_DOWNLOADS]: The skill describes an experimental 'Remote Tasks' feature in references/advanced.md that allows the mise configuration to import and execute task definitions from external URLs (include = ["https://example.com/tasks.toml"]). This creates a vector for loading and executing untrusted code from remote sources.
[DATA_EXFILTRATION]: Documentation in references/bootstrap-monorepo.md and references/env-integration.md directs the agent to access sensitive credential locations, specifically files within the ~/.claude/.secrets/ directory and .env.secrets. Directing an agent to specific sensitive file paths increases the risk of credential exposure and potential exfiltration.
[COMMAND_EXECUTION]: Examples provided in references/arguments.md demonstrate unsafe shell command construction where user-provided arguments are interpolated into commands without proper quoting or validation (e.g., run = 'cat ${usage_files}'). This documentation encourages patterns that are susceptible to shell command injection attacks.

mise-tasks