Architecture-first PR review skill for Claude Code. Clone this repo, open Claude Code, and it auto-guides you through installation into any project.
Installs a Claude Code skill that reviews your pending PRs with focus on:
- Data model scalability
- Genericity and reuse
- Module hierarchy placement
- Contract boundaries and wire formats
- Dependency chains and merge order
- Side-effects and fragile wiring
- Deploy order safety
- Semantic correctness of side-effects
It also installs a deep-review agent that runs the base review, then spawns subagents to investigate each major concern in depth.
git clone https://github.com/TensorOpsAI/dg-review.git
cd dg-review
# Open Claude Code here — it will guide you through setupClaude will ask you:
- Which project to install into (path to your repo)
- Your GitHub username (auto-detected from
gh auth) - How you discover PRs (label, review-request, or both)
- What to name the skill
Then it writes the configured skill directly into your project's .claude/ directory.
- Claude Code installed
ghCLI installed and authenticated- Access to the GitHub repository you want to review PRs for
your-project/
.claude/
skills/<skill-name>/SKILL.md # The review skill (slash command)
agents/deep-review.md # Deepdive agent
In Claude Code, from your project directory:
| Command | What it does |
|---|---|
/<skill-name> |
Full PR review — discovery, ordering, resolved comments, architecture analysis |
deep-review agent |
Runs the review, then spawns subagents for each major architectural concern |
If you prefer to run the shell installer directly:
cd /path/to/your/project
chmod +x /path/to/dg-review/install.sh
/path/to/dg-review/install.shAfter installation, edit the generated SKILL.md in your project to:
- Add/remove architecture checks (Step 6 sections)
- Change the output format (Step 7)
- Adjust what NOT to review
- Tune the tone
| Mode | How it finds PRs |
|---|---|
label |
PRs with a specific GitHub label |
reviewer |
PRs where you're a requested reviewer |
both |
Label AND review-requested combined |
MIT