pr-review-prep

An OpenClaw skill that turns a GitHub pull-request URL into a structured review checklist, with risk flags derived from file-path heuristics rather than model guesswork.

This is the companion skill to the DEV.to post Why This Backend Engineer Stopped Calling LLM APIs From Every Service And Started Running a Local Agent Instead.

Why this exists

Pull-request review is the place where your organisation's risk tolerance shows up in practice. Most "AI PR reviewer" tools fail the same way: they make the LLM decide what is risky, which is the one thing the LLM is actually bad at, because your risk model is not in the training data.

This skill inverts that:

• Deterministic heuristics (a small bash script, ~60 lines) decide which files are risky. Every rule is one line of grep -E you can read in git.
• The LLM composes the human-readable checklist and open questions from those flags, where language-model judgement actually helps.

The split means you can audit the risk logic in a PR, and the skill's behaviour is predictable enough to trust in a team setting.

Install

# 1. Clone next to your other OpenClaw skills
git clone https://github.com/<you>/pr-review-prep.git \
  ~/.openclaw/skills/pr-review-prep

# 2. Ensure prerequisites
gh auth status              # must be authenticated
which bash && which grep    # standard POSIX tools

# 3. Reload skills in the OpenClaw dashboard or run:
openclaw skill reload pr-review-prep

The skill advertises its dependencies in SKILL.md under metadata.openclaw.requires, so OpenClaw will offer to install gh via Homebrew or apt on first use.

Use

In any channel wired to OpenClaw (CLI, Slack, iMessage):

> review prep for https://github.com/openclaw/openclaw/pull/123

The agent pulls the diff with gh, runs scripts/risk-scan.sh against the file list, and returns a markdown checklist with:

• PR header (title, author, branches, +/− line counts)
• Risk flags from the heuristic
• Reviewer checklist (one item per flag + test coverage + rollback)
• Open questions derived from the PR body

See SKILL.md for the full output contract and an example.

Extend

Teams should fork and add their own heuristics. Good candidates:

• Flag changes to cron schedules or retry configuration.
• Flag PRs that touch a "golden file" that only two engineers own.
• Flag PRs that modify the PR-template itself (so the person reviewing the template is aware it will propagate).

All of these are one-line additions to scripts/risk-scan.sh. Open a PR and the rules become part of your team's reviewable policy, not a prompt hiding inside an agent config.

License

MIT. See LICENSE.