Agentic Coding Setup

A personal Claude Code setup for LLM engineering — custom subagents, skills, and a library of reusable components organized by context.

---

Structure

Agentic-Coding-Setup/
├── setup/              # Active setup: copy these into your project or ~/.claude/
│   ├── common/         # Always useful, regardless of project type
│   ├── development/    # For software development projects
│   └── research/       # For ML research and LLM training projects
│
├── library/            # Secondary library: browse and install on demand
│   ├── agents/         # General-purpose agents
│   └── skills/         # Specialized skills
│
└── CLAUDE.md           # Repo-level instructions for Claude Code

Setup vs Library

• setup/ — Opinionated, curated components installed directly into .claude/. Divided by project context.
• library/ — A broader collection for ad-hoc use. Not installed by default. Browse with find-skill or hire-subagent, then install what you need.

---

Setup — Common

Always available, regardless of project type.

Agents help you manage the setup itself: searching and installing from the library on demand.

Skills cover cross-cutting workflows: git conventions, and authoring new agents and skills from scratch.

---

Setup — Development

For software engineering and application development projects.

Agents cover full-stack Python AI application development and codebase exploration.

Skills cover the core AI development stack: LangGraph, vLLM, LangChain, MCP, quantization (AWQ, GPTQ), and workflow utilities.

---

Setup — Research

For ML research, LLM fine-tuning, and experimentation.

Agents are specialized for the full ML research workflow:

• Training: SFT (LoRA variants, Axolotl, Unsloth, TRL) and RL post-training (GRPO, DAPO, DPO, SimPO, verl)
• Evaluation: automated benchmarks, task-specific evals, regression checks
• Data & Literature: dataset preparation, paper reading and analysis

Each agent links to references/ subdirectories with deep domain knowledge — algorithm details, failure modes, config patterns — without bloating the main agent file.

Skills cover the research toolkit: fine-tuning frameworks, evaluation harnesses, quantization libraries, vector search, embedding models, and research methodology (experiment reporting, ablation design, ideation).

---

Library

A broader set of general-purpose agents and specialized skills, available on demand via find-skill or hire-subagent.

• Agents: engineering roles (backend, devops, ML, NLP, security, etc.) and specialist functions (code review, debugging, documentation, performance)
• Skills: ML frameworks, serving infrastructure, document handling, model optimization techniques, and more

Browse the library/ directory or let find-skill search for you.

---

How to Use

Install everything from a category

# Copy research agents and skills into your global Claude config
cp -r setup/research/agents/* ~/.claude/agents/
cp -r setup/research/skills/* ~/.claude/skills/

Browse and install from the library

Ask the find-skill agent in Claude Code:

"Find me a skill for working with FAISS"

Or use hire-subagent:

"I need an agent to review my API design"

Start fresh for a new project

Copy setup/common/ into your project's .claude/, then add whichever of development/ or research/ fits the project.

---

Key Design Principles

• Context-based organization: common / development / research — install only what's relevant
• Setup vs Library: opinionated primary set + broader secondary library for on-demand use
• Reference files for depth: agents and skills link to references/ subdirectories for domain knowledge without bloating the main file
• LIMA principle: quality over quantity — each component should be something you'd actually use