Claude Code Setup for AI Engineers

A portable Claude Code configuration kit built for AI engineers — model training, research, prototyping, and shipping ML systems. One npx command bootstraps a project with the skills, agents, workflows, and plugins you actually use day-to-day.

Not a generic dev kit. The defaults are tuned for ML/research work: PyTorch Lightning, Pydantic, Hugging Face, paper reading and writing (LaTeX/Typst), Python best practices, and end-to-end MR review.

Why use this

• AI/ML out of the box — huggingface-skills, pytorch-lightning, pydantic, python-best-practices, code-refactor configured and ready.
• Research workflow built in — latex-paper-en, latex-thesis-zh, typst-paper, research-paper-writing, paper-audit, industrial-ai-research.
• Reproducible setups — every project gets the same CLAUDE.md, rules, allowlist, and agent definitions.
• Multi-agent orchestration — orchestrator + subagents (feature-developer, test-engineer, code-reviewer, git-manager, ci-validator) that map onto a typical training/prototyping → MR loop.
• GitLab and GitHub — pick one in the installer; CI validation and MR review skills wire up automatically.

Quick start

Requirements: Node.js 18+, Claude Code installed.

From inside your project directory:

npx github:arnabdeypolimi/claude_code_setup

The installer asks for:

1. Target directory
2. VCS platform (GitLab / GitHub / Neither)
3. Skill groups to install (Essential is always included)
4. Project name

If claude is on your PATH, plugins auto-install. Otherwise the exact commands are printed.

To update skills later:

npx github:arnabdeypolimi/claude_code_setup --update

The updater detects changed/added skills, shows a multiselect, and resolves local edit conflicts (backup / overwrite / skip).

Diagnose drift — doctor

After install, your project has a claude-setup.lock.json that records which skills and plugins were set up. doctor checks that nothing has drifted:

npx github:arnabdeypolimi/claude_code_setup doctor

It reports:

• Skills listed in the lock but missing from .claude/skills/
• Skills present on disk but not in the lock
• Plugins enabled in the lock but missing from .claude/settings.json
• Plugins enabled in .claude/settings.json but not in the lock

Exits non-zero if drift is found, so it's safe to run in CI.

AI engineer workflows

Concrete things you can do once installed:

You want to...	Skills / plugins that fire
Fine-tune a model on HF Jobs	huggingface-skills:huggingface-llm-trainer, huggingface-vision-trainer, huggingface-trackio
Pick the right model for a task	huggingface-skills:huggingface-best, huggingface-local-models
Wire training code into Lightning	pytorch-lightning, python-best-practices
Add typed validation to a serving API	pydantic, python-best-practices
Read and summarize a paper	huggingface-skills:huggingface-papers, paper-audit
Write or polish a paper (LaTeX / Typst)	latex-paper-en, latex-thesis-zh, typst-paper, research-paper-writing
Research a domain (e.g. industrial AI)	industrial-ai-research
Run a structured MR review	pr-review-toolkit, code-review, code-simplifier
Refactor legacy Python to current standards	code-refactor
Diagram a system or training pipeline	mermaid-diagrams
Build a Gradio demo	huggingface-skills:huggingface-gradio
Run Transformers in JS / browser	huggingface-skills:transformers-js

What's inside

Type	Location	Purpose
Skills	.claude/skills/	Custom slash-command prompts
Agents	agents/	Orchestrator and subagent definitions
Workflows	workflows/	Multi-step workflow definitions
Templates	templates/	Starter configs to copy into new projects
Settings	.claude/settings.json	Global model config and enabled plugins

Skills

AI / ML / data

Skill	Description
pytorch-lightning	Organize PyTorch into LightningModules, multi-GPU/TPU, DDP/FSDP/DeepSpeed
pydantic	v2 validation for APIs, settings, ORM models — Rust-powered core
python-best-practices	Type-first development with dataclasses, discriminated unions, Protocol
python-docs	Generate and sync API references and Mermaid architecture diagrams
code-refactor	Refactor Python to repo coding standards (best-practices + Pydantic v2 + Lightning)

Plus the full huggingface-skills plugin: huggingface-llm-trainer, huggingface-vision-trainer, huggingface-best, huggingface-local-models, huggingface-datasets, huggingface-trackio, huggingface-gradio, huggingface-papers, huggingface-paper-publisher, huggingface-community-evals, huggingface-tool-builder, transformers-js, hf-cli.

Research and academic

Skill	Description
industrial-ai-research	Literature research for industrial AI (anomaly detection, predictive maintenance, scheduling)
latex-paper-en	English LaTeX paper assistant — compile, grammar, logic, de-AI, translation
latex-thesis-zh	Chinese LaTeX thesis — GB/T 7714, thuthesis/pkuthss, structure review
paper-audit	Multi-agent paper review — critical, domain, literature, methodology reviewers
research-paper-writing	ML/CV/NLP paper writing — intro, abstract, method, experiments guides
typst-paper	Typst academic paper assistant — same coverage as latex-paper-en for Typst

Engineering and ops

Skill	Description
git-operations	Git branch, commit, rebase, stash patterns
gitlab-integration	glab CLI for MRs, issues, CI/CD
mermaid-diagrams	Class, sequence, flow, ER, C4, state diagrams
remotion-best-practices	Video creation in React with Remotion
skill-creator	Guide for creating new skills

UI / UX (impeccable)

Source: pbakaus/impeccable. Useful when shipping demos, dashboards, or evaluation UIs around your models.

frontend-design, animate, adapt, arrange, audit, bolder, clarify, colorize, critique, delight, distill, extract, harden, normalize, onboard, optimize, overdrive, polish, quieter, teach-impeccable, typeset.

Agents

Agent	Model	Role
orchestrator	opus	Coordinates multi-agent workflows, delegates to subagents
feature-developer	opus	Implements features and fixes following project patterns
test-engineer	sonnet	Creates unit, integration, and edge-case tests
code-reviewer	sonnet	Reviews for quality, security, and best practices
git-manager	sonnet	Manages branches, commits, and MRs via glab
ci-validator	sonnet	Troubleshoots GitLab CI pipelines

Workflows

Workflow	Description
feature-workflow	End-to-end feature implementation: requirements → MR
cascade-mr	Cascading MRs where multiple task branches build on each other
mr-review	Automated MR review: fetch diff → analyse → post feedback

Skill groups (installer)

Group	Skills	Plugins
Essential (always)	git-operations	superpowers, context7, claude-md-management, pr-review-toolkit, code-review, code-simplifier, claude-mem
GitLab	gitlab-integration	gitlab
GitHub	—	github
Python Development	python-best-practices, python-docs, pydantic, pytorch-lightning, code-refactor	—
AI / Machine Learning	—	huggingface-skills
Research & Academic	industrial-ai-research, latex-paper-en, latex-thesis-zh, paper-audit, research-paper-writing, typst-paper	academic-research-skills
Frontend / UI-UX	21 impeccable skills	frontend-design, playwright, ui-ux-pro-max
Diagrams & Docs	mermaid-diagrams, skill-creator	—
Video / React	remotion-best-practices	—
Security	—	security-guidance
Marketing	—	marketing-skills
Caveman Mode	—	caveman
Agent Skills	—	agent-skills

Repo layout

claude_code_setup/
├── README.md
├── CLAUDE.md
├── create-claude-setup/        # Installer (npx entry point)
├── .claude/
│   ├── settings.json           # Enabled plugins and global model config
│   └── skills/                 # All skill definitions
├── agents/
│   ├── orchestrator/orchestrator.md
│   └── subagents/{feature-developer,test-engineer,code-reviewer,git-manager,ci-validator}.md
├── workflows/
│   ├── feature-workflow.md
│   ├── cascade-mr.md
│   └── mr-review.md
└── templates/
    ├── settings.json           # Starter project settings.json
    ├── config.yaml             # Master project configuration
    ├── rules.md                # Global rules and conventions
    ├── allowlist.yaml          # Security permissions
    ├── commit-message.md       # Commit message format guide
    ├── mr-template.md          # MR description template
    └── review-template.md      # Code review output template

Manual setup

If you'd rather not run the installer:

1. Clone this repo:

   git clone https://github.com/arnabdeypolimi/claude_code_setup.git

2. Copy templates into your project:

   mkdir -p <your-project>/.claude
   cp claude_code_setup/templates/settings.json <your-project>/.claude/
   cp claude_code_setup/templates/config.yaml <your-project>/.claude/
   cp claude_code_setup/templates/rules.md <your-project>/.claude/
   cp claude_code_setup/templates/allowlist.yaml <your-project>/.claude/

3. Install plugins from inside Claude Code:

   claude plugin install superpowers
   claude plugin install context7
   claude plugin install pr-review-toolkit
   claude plugin install code-simplifier
   claude plugin install claude-md-management
   claude plugin install code-review
   claude plugin install skill-creator
   claude plugin install gitlab
   claude plugin install playwright
   claude plugin install security-guidance
   claude plugin install huggingface-skills
   claude plugin install frontend-design
   claude plugin install firebase

4. Register custom marketplaces (one-time, global):

   claude plugin marketplace add nextlevelbuilder/ui-ux-pro-max-skill
   claude plugin marketplace add thedotmack/claude-mem
   claude plugin marketplace add coreyhaines31/marketingskills
   claude plugin marketplace add JuliusBrussee/caveman
   claude plugin marketplace add addyosmani/agent-skills
   claude plugin marketplace add Imbad0202/academic-research-skills

5. Install marketplace plugins:

   claude plugin install ui-ux-pro-max@ui-ux-pro-max-skill
   claude plugin install claude-mem@thedotmack
   claude plugin install marketing-skills@marketingskills
   claude plugin install caveman@caveman
   claude plugin install agent-skills@addy-agent-skills
   claude plugin install academic-research-skills@academic-research-skills

Plugin marketplaces

ID	Source	Description
claude-plugins-official	anthropics/claude-plugins-official	Official Anthropic plugins
ui-ux-pro-max-skill	nextlevelbuilder/ui-ux-pro-max-skill	UI/UX design intelligence
thedotmack	thedotmack/claude-mem	claude-mem (cross-session memory)
marketingskills	coreyhaines31/marketingskills	Marketing skills (CRO, copywriting)
caveman	JuliusBrussee/caveman	Ultra-compressed responses
addy-agent-skills	addyosmani/agent-skills	Addy Osmani's agent skills
academic-research-skills	Imbad0202/academic-research-skills	Academic research pipeline (research → write → review → revise → finalize)

Global settings

~/.claude/settings.json is applied across all projects. The owner's default:

{ "model": "sonnet" }

Project-level .claude/settings.json overrides this per repo. settings.local.json is machine-specific and is not committed.