The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill uses bunx or npx to execute the skills CLI tool, which fetches and applies instructions from remote repositories. This allows for the dynamic execution of instructions from arbitrary, user-controlled sources, potentially leading to unauthorized operations.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to download a remote SKILL.md file and incorporate its instructions into the current conversation. This creates a surface for indirect prompt injection if the remote repository contains malicious content.
Ingestion points: Remote SKILL.md files fetched from sources like GitHub via the skills tool (documented in SKILL.md).
Boundary markers: The instructions include manual guardrails, telling the agent to treat fetched instructions as task guidance and not as permission for sensitive actions like accessing credentials.
Capability inventory: The agent is directed to carry out the instructions found in the remote file, potentially leveraging any tools available to it.
Sanitization: The skill lacks automated sanitization or verification of the remote markdown instructions before they are integrated into the prompt.
[COMMAND_EXECUTION]: The skill directs the agent to execute shell commands, including git rev-parse --show-toplevel and the assembly of command lines for bunx and npx using user-supplied parameters. It also manages local configuration files at ~/.config/use-remote-skill/config.yaml and .agents/remote-skills.yaml.
[PERSISTENCE]: The skill includes functionality to modify AGENTS.md files to inject persistent rules that cause the agent to automatically search and load remote skills from the catalog whenever a new task begins.

use-remote-skill