The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill provides instructions for AI agents to perform self-serve installations of CLI tools when missing, specifying that these should come from trusted official repositories and use privilege-free, user-directory paths.
Evidence: implementation-workflow.md describes the self-serve installation process and README.md provides git clone commands for the skill itself, targeting the author's repository.
[COMMAND_EXECUTION]: The skill governs the design and invocation of CLI tools by AI agents. It emphasizes using absolute paths and non-interactive modes to ensure safe and predictable command execution.
Evidence: SKILL.md and references/design-principles.md detail the requirements for exit codes, non-interactive flags, and execution verification.
[INDIRECT_PROMPT_INJECTION]: The skill specifically addresses the security risks of agents processing external data from tools and provides guidance on robust input validation and structured output handling to prevent injection attacks.
Ingestion points: Data ingestion occurs through the execution of external CLI tools and processing their stdout (JSON or text).
Boundary markers: The skill mandates the use of structured JSON envelopes ({ok, data, error, meta}) to clearly delineate success and failure states.
Capability inventory: The guidelines cover file writes, network operations (via the CLI being designed), and local command execution.
Sanitization: references/design-principles.md contains a specific section on hard-blocking path traversal and command injection, and references/pitfalls.md warns against common implementation mistakes.
[SAFE]: The skill focuses on promoting security best practices, such as optimistic locking via fingerprints for updates/deletes, secure credential storage (0700/0600 permissions), and avoiding persistence in shared environments. All provided URLs and package names are consistent with the author's identity.

agent-friendly-cli