The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from GitHub pull request comments, review summaries, and CI logs, creating a surface for indirect prompt injection.
Ingestion points: Data is ingested via GraphQL queries in shared/pull-request-comments.md (conversation comments, review bodies, and inline threads) and CI log retrieval in SKILL.md (gh run view --job <id> --log-failed).
Boundary markers: The skill explicitly references shared/external-data.md, which mandates labeling external text as untrusted and fencing it with backticks to prevent the agent from following embedded instructions.
Capability inventory: The agent has the ability to execute shell commands (gh, git), write files, and commit/push code changes to the user's repository.
Sanitization: Instructions in shared/external-data.md forbid interpolating untrusted prose into shell command text, requiring the use of temporary files (--body-file) or environment variables instead.
[COMMAND_EXECUTION]: The skill performs extensive local and remote operations using the GitHub CLI (gh) and Git.
Evidence: Commands include gh pr view, gh pr checks, gh pr ready, git add, git commit, and git push.
Security measures: The skill implements 'Branch binding' in references/ci-fix.md to ensure that local Git actions correctly target the intended upstream PR branch and prevent unintended pushes to other remotes or branches.
[EXTERNAL_DOWNLOADS]: The skill includes instructions for the user to install a dependency using a package manager.
Evidence: SKILL.md directs users to run npx skills@latest add bostonaholic/skills --skill addressing-pr-comments if the required sub-skill is missing. This refers to the author's own tooling ecosystem.

watching-authored-prs