Skip to content

ai-usage-guidelines.md

Latest commit

 

History

History
23 lines (12 loc) · 2.04 KB

ai-usage-guidelines.md

File metadata and controls

23 lines (12 loc) · 2.04 KB

Guidelines for AI-Assisted Development

Guiding Principle

A human must ultimately take responsibility for all aspects of library code.

Project Policies

  1. Human Manager Responsibility: The human manager is responsible for reviewing and fully understanding all contributed code.

  2. Commit Attribution: Commits may be attributed to an agentic AI tool, but all commit messages should include the @username of the human manager who is ultimately responsible for them. For small features, consider squash merges to condense work into a single joint-authored commit to indicate the human associated with any AI contributions.

  3. Pull Request Review Process: Pull requests involving LLM-assisted code generation or attributed to an agentic AI tool should go through the usual peer review process, with the human manager responsible for the PR completing a pull request code review process first before requesting peer review.

  4. Effort Balance: The effort contributed by the human manager responsible for LLM-generated code and commits should exceed the effort requested from peer reviewers.

  5. Limits on extent of Agentic AI use: Agentic AI may be used to perform simple maintenance tasks (e.g., figuring out we need to update a version number) but not to generate new core-library source code.

  6. Energy Safeguards for Agentic AI use: Agentic AI should not be run automatically without first reaching a consensus among primary developers, and there should be restrictions added to ensure it stops promptly if it is not able to solve a problem.

  7. In-code citations: Any code that was added via an AI should be clearly marked and ideally include a link to the chat session that generated it.

  8. Standard of oversight: Contributors and reviewers are expected to have mentally stepped through the code and made sense of it before requesting or approving a review. Exemptions should be requested explicitly where the judged effort tradeoff is not worthwhile. This standard will be subject to further consideration and possible future adjustment.