A collection of reusable Claude Code skills and prompts for common development workflows.
This library reflects explorations with high and low levels of delegation.
One thing is certain: first pass LLM output is almost always unsatisfactory. Iteration helps, but it's not a magic wand or a replacement for critical thinking.
review-pr |
review-pr-deep |
review-pr-escalate |
|
|---|---|---|---|
| Passes | 1 (orchestrator does everything) | 1 round, 3 parallel specialists | 5 sequential passes, up to 3 iterations each |
| Agents | None (one conversation) | General + security + integration (by PR type) | Fresh agent per pass/iteration |
| Lens | Design + full code checklist in one go | Domain split (quality vs security vs wiring) | Zoom out over time (code → gaps → design → architecture → convergence) |
| Fixes code? | No | No (optional converge.md loop) |
Yes (auto-fixes issue/todo between passes) |
| Prior reviews | No | Yes (fetches existing comments/reviews) | No |
| Best for | Quick, efficient review | Non-trivial PRs needing specialist coverage | Critical PRs you want reviewed until converged |
Single-pass PR review with ruthless critical rigor. Provides structured, clearly labeled findings for design, correctness, tests, clarity, and style.
Use when: You need a quick, efficient review or one pass is sufficient.
/review-pr <PR-number>
/review-pr <branch-name>Multi-agent PR review with specialized passes: general code review, integration testing, and security analysis. Agents converge on findings.
Use when: You need a thorough, comprehensive review of high-stakes or complex changes.
/review-pr-deep <PR-number>Escalated multi-pass review covering code quality, design gaps, architecture alignment, and final convergence. For the most critical reviews.
Use when: You need exhaustive analysis across five specialized passes.
/review-pr-escalate <PR-number>Expert guidance on Python best practices, style guides, type hints, and code quality. Knowledge base includes PEP 8, Google Style Guide, PEP 257 (docstrings), and PEP 484 (type hints).
Use when: You need advice on Python coding standards, naming conventions, or documentation.
/python-expert How should I name my class variables?
/python-expert What's the correct format for multi-line docstrings?
/python-expert Should I use List[int] or list[int]?Monitors a PR for CI failures and bot review comments. Diagnoses issues, applies fixes, and pushes. Repeats until all checks pass.
Use when: You want automated CI fixing until a PR is green.
/babysit-ci
/babysit-ci <PR-number>
/babysit-ci <branch-name>Works well with /loop for hands-off monitoring:
/loop 1m /babysit-ci 359These skills integrate directly with Claude Code. To use them:
- Ensure Claude Code is installed. Available as CLI, desktop app, web app, or IDE extensions.
- Authenticate with git and gh. Required for all PR and CI-related skills:
git config user.name "Your Name" git config user.email "you@example.com" gh auth login
- Navigate to a git repository — Skills operate on the current repo's code and PRs
- Invoke a skill using
/skill-namein Claude Code (e.g.,/review-pr 123).
Each skill is self-contained with:
- SKILL.md — Detailed step-by-step instructions and methodology
- knowledge/ — Reference materials (Python expert knowledge base)
- agents/ — Specialized agent definitions (review-pr-deep, review-pr-escalate)
- Be firm on principles, flexible on implementation. Require that a problem be solved while giving the author latitude in how to solve it.
- Be ruthlessly direct. State what is wrong and why. Don't hedge or soften.
- Label every finding using Conventional Comments format.
- Provide concrete alternatives when proposing changes.
- Code is read more often than written. Prioritize readability.
- Consistency matters. Consistency within a module matters more than within a project, which matters more than with style guides.
- Explicit is better than implicit. Write clear code over clever code.
- Minimal fixes only. Fix exactly what CI is complaining about, don't refactor
- Understand before fixing. Read the error and code, then act.
- Report what you did. Explain the diagnosis and fix for user visibility.
# Quick review
/review-pr 123
# Comprehensive multi-agent review
/review-pr-deep 123
# Exhaustive five-pass review
/review-pr-escalate 123/python-expert How do I write a proper docstring for a function with multiple return values?
/python-expert What are the differences between Google Style and PEP 8 naming conventions?# One-shot fix
/babysit-ci 359
# Hands-off loop until green
/loop 1m /babysit-ci 359- Git CLI and gh CLI (GitHub CLI) installed and authenticated
- Current working directory in a git repository (for most skills)
- Push access to branches (for babysit-ci)
Created by Anna Baker (@annabkr)