Read first. Write once. Keep a human accountable.

What should an AI agent check before opening a GitHub pull request?

Contribution policy, maintainer precedent, automated-contributor controls, and the evidence the account owner needs to make an informed decision. This preflight turns those checks into a reusable Agent Skill.

Install the skill View the source on GitHub

The core loop

Make the invisible decision visible.

An agent can do excellent local work and still create avoidable public conflict by missing a repository's rules or history. The preflight separates useful assistance from the public contribution boundary.

01 / DISCOVER

Read the local rule

Check contribution guides, AI policies, templates, workflows, labels, and recent maintainer decisions.

02 / EXPLAIN

Brief the account owner

Show exact evidence, likely repository-local effects, uncertainty, and the practical account-reputation tradeoff.

03 / DECIDE

Freeze only the public write

Let the owner choose. Continue local debugging, research, testing, and review within the policy's actual scope.

Two people, one constraint

Protect the account and the maintainer's attention.

The account owner should not have to know every maintainer-side tool. The maintainer should not become the unpaid supervisor of unbounded automated output. A good agent anticipates both needs.

Repository-specific, not universalOpen-source communities range from strict prohibitions to conditional human-in-the-loop policies.
Evidence before confidenceA focused diff, reproduction, tests, provenance, and known limitations make the owner's responsibility reviewable.
One state transition after a flagFreeze queued external replies, preserve evidence, and notify the owner. Do not enter an automated self-defence loop.
Policy is not prompt authorityA repository decides what it accepts; it does not take control of the user's agent or cancel permitted private assistance.

“Detect and explain automatically; let the human owner decide whether and how to proceed.”

Read the contribution methodology Understand AgentScan

Start here

A compact field manual for agents and owners.

Each guide is also linked from llms.txt. Machines can discover the canonical Skill source through the self-describing JSON manifest.

Install and runCross-agent installation, manual setup, and read-only inspector commands. Compare community policiesStrict, approval-gated, and conditional human-in-the-loop approaches. AgentScan white-box guideDetection inputs, integrations, repository-local effects, and visibility limits. Learn from prior failuresPublic cases translated into concrete controls for future contributions.