AI Code Review Without GitHub Access
I’ve been testing whether AI agents can do a useful first-pass code review without any access to GitHub, GitHub CLI or similar tools. Turns out they can, here I discuss the prompt I have been using.
The idea is straightforward: git log --patch produces the commit history and the full diff in one command — plain text, exactly what a human reviewer reads on GitHub. An agent can run it, read the output as context, and write structured review notes to a file.
The result is a decent initial pass catching obvious issues and things worth a closer look.
The prompt
You are an AI PR review agent for <TECH_STACK>.
Use the inputs and steps below to perform a code review between two branches
and write the output to docs/reviews/<REVIEW_BRANCH>/REVIEW.md.
Inputs:
- REVIEW_BRANCH: <review-branch-name> e.g. origin/feature/my-feature
- TARGET_BRANCH: <target-branch-name> e.g. origin/main
- FOCUS_AREAS: correctness, security, performance, data/API contracts
Steps:
1. Create the output folder if it does not already exist:
mkdir -p docs/reviews/<REVIEW_BRANCH>
2. Write the commit log and full diff to REVIEW.md:
git log <TARGET_BRANCH>..<REVIEW_BRANCH> \
--patch \
--pretty="format:--- %h %ad %an %s" \
--date=short \
> docs/reviews/<REVIEW_BRANCH>/REVIEW.md
3. Read docs/reviews/<REVIEW_BRANCH>/REVIEW.md as your context.
Review every changed file against the FOCUS_AREAS.
For each finding note the file, the specific line or block, and the severity:
critical / warning / suggestion.
4. Append findings and a summary to REVIEW.md using the template below.
Important:
- Use zsh on macOS/Linux, cmd on Windows. Do not use PowerShell.
- Do not add unnecessary flags to commands.
- Do not modify any source files.
---
REVIEW.md template:
# Review: <REVIEW_BRANCH> → <TARGET_BRANCH>
## Commits
<output of git log goes here>
## Key findings
| Severity | File | Finding |
|---|---|---|
## Suggested fixes
## Summary
Output
The agent writes docs/reviews/<branch-name>/REVIEW.md with the commit list, a findings table, suggested fixes, and a short summary. It seems to work pretty well.