The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes several local shell scripts for pipeline management and repo hygiene, including pre-ship-check.sh, quality-check.sh, resolve.sh, and save-artifact.sh. It also leverages standard development tools like the GitHub CLI (gh), git, and jq for session state management.
[DYNAMIC_EXECUTION]: The skill sources external shell scripts from local paths at runtime, such as bin/lib/git-context.sh and ~/.claude/skills/nanostack/bin/lib/skill-preamble.sh. These scripts are part of the internal framework provided by the skill author.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted local data (branch names, commit messages, and file contents) and interpolates them into PR previews and templates.
Ingestion points: Untrusted data enters the agent context through git status, branch name detection, and file reading during pre-flight checks in SKILL.md and the bin/ scripts.
Boundary markers: A mandatory PR Preview step is implemented in SKILL.md section 2, requiring explicit user approval before the data is committed to a public PR.
Capability inventory: The skill has capabilities to create PRs (gh pr create), push code (git push), and execute shell scripts via the PreToolUse hook.
Sanitization: There is no automated escaping of interpolated content; verification relies on the user during the mandatory preview stop.

ship