The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill analyzes data from pull requests and chat transcripts to infer user preferences, which represents a potential surface for instructions embedded in external content.
Ingestion points: Data is read from the GitHub API using scripts/collect_pr_feedback.py and from local transcript files via scripts/collect_chat_feedback.py.
Boundary markers: Explicit instructions in SKILL.md and references/history.md warn the agent to treat this content as untrusted evidence rather than direct instructions.
Capability inventory: The skill can update project instruction files (AGENTS.md, CLAUDE.md) and create symlinks using scripts/connect_project.py. It also performs network requests and data retrieval via the GitHub CLI in scripts/collect_pr_feedback.py and writes temporary data in scripts/history_common.py.
Sanitization: Results are filtered for supported message types and schemas. All proposed criteria updates require explicit user review and confirmation before being persisted to the ALIGNS.md file.
[COMMAND_EXECUTION]: The skill uses the GitHub CLI (gh) to retrieve project metadata and feedback history.
Evidence: scripts/collect_pr_feedback.py executes gh api via subprocess.run to fetch pull request information. The command is invoked with a hardcoded argument list and the repository name is validated against a strict regex to prevent shell injection.
[EXTERNAL_DOWNLOADS]: The skill references external assets and leverages established machine learning libraries for local analysis.
Evidence: references/badge.md includes a link to an SVG asset hosted on the author's official GitHub repository. scripts/check_training_adapter.py utilizes the transformers library to perform tokenization checks; notably, it is configured with trust_remote_code=False to prevent the execution of arbitrary remote code during the check.

align