liveBook · Manning

Instruction file

As explained in AI Governance

An agent instruction file contains directions that govern how an AI coding agent behaves in a codebase.

An agent instruction file tells an AI coding agent how to behave in a codebase, including which tools to run, which conventions to follow, and when human approval is required. A malicious change can silently alter behavior across the development workflow. section 4.5.1

Treat instruction files as security-sensitive artifacts:

  • Include changes in code review and require approval from a security-aware reviewer before they take effect. section 4.5.1
  • Use CODEOWNERS or an equivalent to require internal sign-off on changes, especially in repositories that accept outside contributions. section 4.5.1
  • Flag non-printable Unicode in configuration files; sanitize files from external sources before use. section 4.5.1

How AI Governance covers it

How deep it goes
Prescribes review, approval, Unicode scanning, and sanitization controls without detailing instruction-file syntax.
Examples
Names AGENTS.md, .cursorrules, and .github/copilot-instructions.md, and flags zero-width Unicode characters as a risk.
What you can do after
Operate instruction-file controls through security-aware review, ownership rules, Unicode detection, and sanitization.
How it connects
Connects AI coding agents with code review, CODEOWNERS, CI pipelines, and prompt injection.

Also explained in

Compare how 2 books explain instruction file →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of AI Governance. The section links point to the parts of the book it draws on.