Agent working rules

This is an optional starting point. Keep repository-specific conventions in the repository that owns them.

Use relevant skills

• Check the installed skill list before inventing a workflow.
• Load a matching skill when the request names it or clearly falls within its description.
• Follow repository-local instructions when they are more specific than generic guidance.
• Claude Code can preload skills declared by an agent definition. Codex dispatch must explicitly tell each subagent which SKILL.md files to read completely before acting.

Preserve repository state

• Inspect the branch, worktree, and diff before changing or staging files.
• Preserve unrelated changes and generated local state.
• Stage only reviewed files that belong to the current task.

Protect release boundaries

• Do not push directly to a protected branch unless the current request explicitly names that branch and asks for the direct push.
• Do not merge a pull request unless the current request explicitly names the target branch and asks for the merge.
• Confirm before publishing, deploying, releasing, or creating tags.

Comment sparingly

• Default to zero comments. Code and tests carry the meaning; a comment earns its place only by stating a non-obvious why.
• Never write comments that restate the code, reference a ticket or PR, note provenance, or narrate a change.
• See the architect-review skill for the full guidance.

Report concisely

• Lead with the outcome. Add detail only when it changes what the reader does next.
• Skip the preamble, the recap of the request, and the closing summary of what was just said.
• Prefer a short list to a paragraph of narrative.

Report evidence

• Run checks appropriate to the change and report their actual results.
• State failures, skipped checks, and remaining risks plainly.