The Agent Skills Directory

[PROMPT_INJECTION]: The skill instructions explicitly suspend standard interactive safety and alignment protocols, such as 'confirm-before-commit' and 'ask about preferences,' to maintain an autonomous workflow.
Evidence: 'It suspends the ask-first / confirm-alignment rules for the duration of the run', 'For the duration of this run, these standing rules are suspended'.
[COMMAND_EXECUTION]: The skill enables the agent to autonomously execute shell commands for version control and repository management without user oversight.
Evidence: Authorization to use git commit, git push, and gh pr create tools as part of the automated shipping sequence.
[DATA_EXFILTRATION]: The skill facilitates the transfer of local repository content to remote servers through standard git operations.
Evidence: The skill is designed to 'commit, push and open a ready PR' as a final step, which involves sending code over the network to GitHub.
[PROMPT_INJECTION]: The skill possesses an attack surface for indirect prompt injection as it processes data from the repository during autonomous execution.
Ingestion points: The skill reads project files, git diffs, and tool outputs using tools like Read, Grep, and Agent (SKILL.md, guide.en.md).
Boundary markers: Briefs provided to subagents act as delimiters by specifying file scopes and prohibiting task expansion.
Capability inventory: The skill has the authority to write to the filesystem and push to remote repositories.
Sanitization: Includes a 'verification gate' that uses top-tier models for a security pass and repository-defined checks to validate work before shipping.

autopilot