Skip to content

ivogabriel19/git-github-workflow-skill

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

1 Commit
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Git & GitHub Workflow Skill for Claude Code

A single SKILL.md (plus a CLAUDE.md drop-in and bundled CI templates) that makes Claude Code handle Git, GitHub, and CI the way a senior engineer at a top-tier product company would.

Stack-agnostic. Applies to any language, any project, any team size.

The Problem

LLM coding agents — Claude Code included — tend to handle version control like an afterthought:

  • Commit messages like "updates", "fix stuff", "WIP".
  • One commit mixing a refactor, a feature, and a stray formatting change.
  • 2000-line pull requests that mash three unrelated changes together.
  • Direct pushes to main "just for a quick fix."
  • Blind git push --force that overwrites teammates' work.
  • .github/workflows/*.yml files that ignore caching, pin nothing, and grow into mega-workflows.
  • Merging PRs with red CI because "the fix is obvious."

Each of these is small in isolation. Together they produce a repository whose history is unreadable, whose main is sometimes broken, and whose CI no one trusts.

The Solution

A single opinionated skill, ~440 lines, readable in five minutes. It covers:

Section What it gives you
Pre-flight The three commands to run before any work touches a repo.
Commits Conventional Commits format, atomic-commit rule, staging discipline.
Branches Naming, GitHub Flow, when to rebase, when not to force-push.
Pull Requests When to open, title format, body template, review etiquette, merge strategy.
Branch protection Required settings on main for any serious repo.
CI Stack-agnostic principles + the four job pillars + performance rules.
History rewriting The local-vs-shared rule, amend, interactive rebase, reset.
Tags & releases Semver mapped to commit types, when to tag, changelog generation.
Anti-patterns Cheat-sheet of things to never do, with the right alternative.
Operating defaults A ten-step checklist Claude follows in every session.

The skill is bundled with ready-to-copy templates in skills/git-github-workflow/assets/: a four-pillar ci.yml, a tag-driven release.yml, a PULL_REQUEST_TEMPLATE.md, and a CODEOWNERS skeleton.

Install

Option A — Claude Code plugin (recommended)

Inside Claude Code, add the marketplace and install the plugin:

/plugin marketplace add ivogabriel19/git-github-workflow-skill
/plugin install git-github-workflow@git-skills

The skill becomes available across all your projects, with no per-repo setup.

Option B — CLAUDE.md (per-project)

For a single project, drop the standalone CLAUDE.md into the repo root:

New project:

curl -o CLAUDE.md https://raw.githubusercontent.com/ivogabriel19/git-github-workflow-skill/main/CLAUDE.md

Existing project (append to the existing CLAUDE.md):

echo "" >> CLAUDE.md
curl https://raw.githubusercontent.com/ivogabriel19/git-github-workflow-skill/main/CLAUDE.md >> CLAUDE.md

Bootstrapping a repo with the bundled templates

Once the plugin is installed, ask Claude something like:

Bootstrap this repo with the git-github-workflow conventions. It's a [Python | Node | Go | Rust | …] project.

Claude reads the skill, copies the relevant templates from assets/ into .github/, fills in toolchain placeholders for your stack, and leaves you a PR with the setup.

Repository structure

.
├── .claude-plugin/
│   ├── plugin.json                          # Claude Code plugin manifest
│   └── marketplace.json                     # marketplace entry for /plugin install
├── skills/
│   └── git-github-workflow/
│       ├── SKILL.md                         # the skill itself
│       └── assets/
│           ├── workflows/
│           │   ├── ci.yml                   # four-pillar CI template
│           │   └── release.yml              # tag-driven release template
│           ├── PULL_REQUEST_TEMPLATE.md
│           └── CODEOWNERS
├── CLAUDE.md                                # standalone version for per-project drop-in
├── EXAMPLES.md                              # before/after concrete examples
├── README.md
└── LICENSE

Core principles

The skill in one breath:

  1. main is sacred. Always deployable, always green, never force-pushed.
  2. One commit = one logical change, revertible in isolation without breaking anything.
  3. One branch = one unit of work. Open small, review fast, merge, delete.
  4. Rewrite local history freely. Never rewrite history others have pulled.

The four-pillar CI rule:

Every project runs lint, typecheck, test, and build on every PR and every push to main. All four are required checks. Lint and typecheck run in parallel; build depends on the rest passing.

How to know it's working

You'll see:

  • Commit history that reads like a changelog. Types and scopes are consistent; messages are imperative and bounded.
  • Pull requests under 500 lines that any reviewer can grasp in 30 seconds from the title alone.
  • A main that's always green. No "we'll fix CI on Monday" debt.
  • CI runs under 10 minutes with cached dependencies and concurrency control.
  • git bisect actually works because every commit is atomic.

See EXAMPLES.md for concrete before/after scenarios.

Customization

The skill is opinionated by design but mergeable with project-specific rules. To add overrides, append a section to your project's CLAUDE.md:

## Project-specific workflow overrides

- We use `release-please` to manage tags — Claude should not tag manually.
- All migrations to `db/` require a manual approval from @db-team — add it to the PR checklist.
- Our `main` branch is `master`. Treat `master` everywhere this skill says `main`.

Tradeoff note

The skill biases toward rigor over speed. For trivial repos (personal scratch projects, single-author throwaways), some rules are overkill — use judgment. The goal is to eliminate avoidable mistakes on non-trivial work, not to slow down two-file experiments.

Contributing

Found a workflow case the skill handles badly? Open an issue with the failing prompt and the output you got. PRs welcome — and yes, they need to follow the skill's own conventions. Eat your own dog food.

License

MIT

Packages

 
 
 

Contributors