Personal AI development abstraction layer built on top of Claude Code CLI.
Extend Claude Code with your own skills, plugins, slash commands, and hooks — all outside the CLI package so they survive every upstream update. Works with Anthropic Pro (cloud) or a local Ollama model (offline).
| Requirement | How to check | Required? |
|---|---|---|
| Claude Code CLI | claude --version |
Yes |
| Python 3.10+ | python --version |
Yes |
| Git | git --version |
Yes |
| Ollama | ollama list |
Only for local models |
npm install -g @anthropic-ai/claude-codeNeed Node.js? Get it from nodejs.org. Full install docs: docs.anthropic.com/en/docs/claude-code
If you already have claude working, you can start using this repo's extensions immediately:
git clone https://github.com/Multimodal-Agents/claude-code-building-env.git
cd claude-code-building-env
.\setup.ps1 # clones Anthropic reference repos (skills + claude-code examples)
python -m venv venv
.\venv\Scripts\Activate.ps1
pip install pandas pyarrow numpy requests ddgs huggingface_hubThen just launch Claude from this directory:
claude # uses Anthropic Pro — CLAUDE.md, skills, and slash commands load automaticallyThat's it. Claude Code reads CLAUDE.md and discovers the skills/commands in this repo on startup. No Ollama required.
Want the interactive launcher instead? (lets you pick Anthropic Pro, Ollama, or Speech mode)
.\scripts_and_skills\claude_scripts\run_claude.ps1git clone https://github.com/Multimodal-Agents/claude-code-building-env.git
cd claude-code-building-env.\setup.ps1This clones two Anthropic reference repos into their expected locations:
scripts_and_skills/claude_skills/skills/← anthropics/skillsclaude_code_custom/claude-code/← anthropics/claude-code
If you already have claude-code cloned elsewhere:
.\setup.ps1 -ClaudeCodeDir "C:\your\existing\claude-code"Skip either clone:
.\setup.ps1 -SkipClaudeCode # skip claude-code ref
.\setup.ps1 -SkipSkills # skip skills refpython -m venv venv
.\venv\Scripts\Activate.ps1
pip install pandas pyarrow numpy requests ddgs huggingface_hub.\scripts_and_skills\claude_scripts\run_claude.ps1Or point your desktop shortcut to:
powershell.exe -ExecutionPolicy Bypass -File "M:\claude_code_building_env\scripts_and_skills\claude_scripts\run_claude.ps1"
See QUICKSTART.md for the full step-by-step guide including seeding the CoreCoder dataset and creating the corecoder:latest model.
.
├── CLAUDE.md # Global context injected at every session
├── QUICKSTART.md # Step-by-step setup guide
├── TESTING.md # Test procedures
├── setup.ps1 # One-time repo setup
├── setup_llamacpp.ps1 # Optional: llama.cpp build + Python tools
├── show_dataset.py # Quick dataset viewer
├── assets/ # Icons, static media, video scripts
├── .claude/
│ └── commands/ # 11 slash commands (see table below)
├── scripts_and_skills/
│ ├── claude_scripts/
│ │ ├── run_claude.ps1 # Windows launcher (interactive menu)
│ │ └── run_claude.sh # WSL/Linux launcher
│ ├── data/ # Data layer
│ │ ├── prompt_store.py # Parquet-backed prompt + conversation CRUD
│ │ ├── embeddings.py # nomic-embed-text semantic search
│ │ ├── dataset_generator.py # Docs/code → ShareGPT training conversations
│ │ ├── web_search.py # DuckDuckGo search wrapper
│ │ ├── arxiv_crawler.py # ArXiv Atom API search + paper text fetching
│ │ ├── classifier.py # Content moderation via granite3-guardian
│ │ ├── science_validator.py # Scientific claim validation
│ │ └── seeds/
│ │ └── corecoder_prompts.py # Seeds 19-row CoreCoder prompt dataset
│ ├── model_manager/ # Ollama model lifecycle
│ │ ├── ollama_api.py # REST wrapper: list/create/delete/pull/embed
│ │ ├── modelfile.py # Modelfile builder (system prompt, params, adapters)
│ │ ├── gguf_manager.py # GGUF registry + HF→GGUF conversion + quantize
│ │ ├── lora_manager.py # LoRA adapter registry + Ollama deployment
│ │ └── hf_download.py # Download GGUFs from HuggingFace Hub
│ ├── speech/ # Voice pipeline
│ │ ├── voice_pipeline.py # Real-time speech-to-speech loop
│ │ ├── stt.py # Whisper speech-to-text
│ │ ├── tts.py # edge-tts text-to-speech
│ │ ├── script_narrator.py # Script → narrated audio
│ │ └── audio_utils.py # Audio helpers
│ ├── code_graph/ # Live code graph visualization
│ │ ├── server.py # FastAPI graph server
│ │ ├── code_parser.py # Import/dependency parser
│ │ ├── hook_notify.py # Claude hook listener
│ │ ├── index.html # Cytoscape.js terminal UI
│ │ ├── start.ps1 # Windows launcher
│ │ └── start.sh # WSL launcher
│ └── claude_skills/
│ ├── skills/ # Official Anthropic skills (reference, git-ignored)
│ ├── blues_skills/ # 6 custom domain skills (see table below)
│ └── template_skills/ # Scaffold for new skills
├── agent-chef-claude-code/ # Dataset generation agent
│ ├── ragchef.py # RAG-powered conversation chef
│ ├── conversation_generator.py # Conversation synthesis
│ ├── crawlers_module.py # Web/doc crawlers
│ ├── dataset_expander.py # Dataset augmentation
│ └── classification.py # Quality classification
├── dataset_kitchen/ # Test datasets and samples
├── local_data/ # Runtime data (git-ignored, machine-local)
│ ├── prompts/ # Prompt + conversation parquet datasets
│ ├── embeddings/ # Embedding vectors for semantic search
│ ├── exports/ # Exported datasets
│ └── modelfiles/ # Saved Modelfile blueprints
├── claude_custom_projects_1/ # Active projects (git-ignored)
└── claude_code_custom/
└── claude-code/ # Upstream reference clone (git-ignored)
└── plugins/ # Official plugin examples
Excluded from git (see .gitignore):
venv/— Python virtual environmentlocal_data/— runtime parquet + embedding filesclaude_custom_projects_1/— active work productgit_clones/— third-party clonesclaude_code_custom/claude-code/— upstream referencescripts_and_skills/claude_skills/skills/— Anthropic skills reference clonebasic_reference_documentation_library/— large reference dumps
This repo uses the official Playground plugin for interactive HTML/JS prototyping. The cytoscape-playground skill (below) extends it with Cytoscape.js graph visualization.
Reference plugins from the upstream claude-code repo are cloned into claude_code_custom/claude-code/plugins/ by setup.ps1 for study — they include examples like code-review, feature-dev, hookify, pr-review-toolkit, and more.
| Skill | Path | What it does |
|---|---|---|
| prompt-manager | blues_skills/skills/prompt-manager/ |
Store/search/export prompts and conversations via PromptStore + EmbeddingStore |
| code-parser | blues_skills/skills/code-parser/ |
Surgical codebase reading protocol — single tree scan, mental map, targeted reads only |
| speech-to-speech | blues_skills/skills/speech-to-speech/ |
Real-time voice pipeline: Mic → VAD → Whisper → LLM → edge-tts → speaker |
| cytoscape-playground | blues_skills/skills/cytoscape-playground/ |
Extends Playground plugin with Cytoscape.js graph/network visualization |
| job-list | blues_skills/skills/job-list/ |
Sequential project build queue — processes *_1.md → *_2.md in order, gates on ≥95% completion |
All skills live under scripts_and_skills/claude_skills/blues_skills/skills/.
Available in any claude session launched from this repo:
| Command | What it does |
|---|---|
/prompt-list |
Browse, search, and export the prompt dataset |
/set-system |
Change any Ollama model's system prompt interactively |
/generate-dataset |
Convert files/dirs into ShareGPT training conversations |
/show-dataset |
View rows from any saved dataset (zero-token, raw output) |
/parse-codebase |
Smart codebase exploration with minimal token usage |
/web-search |
DuckDuckGo search (useful for local model sessions) |
/model-manager |
Manage models, GGUFs, Modelfiles, and LoRA adapters |
/script-to-audio |
Parse a video script, strip director notes, render to .wav via edge-tts |
/speech |
Start real-time voice-to-voice mode (say "goodbye claude" to exit) |
/code-graph |
Live terminal CRT code graph — parse imports, watch edits, animate changes |
/job-list |
Run a sequential queue of project builds from numbered markdown specs |
The Claude Code CLI is the engine. This repo is the configuration and extension layer.
[Desktop Shortcut / Terminal]
│
▼
[run_claude.ps1 / run_claude.sh] ← interactive menu: Anthropic Pro, Ollama, or Speech
│
├─── Anthropic Pro ──► claude (cloud, no local model needed)
│
└─── Ollama ────────► claude --model gpt-oss:20b (local, offline)
│
▼
[claude CLI] ← reads CLAUDE.md, loads skills & commands automatically
Benefits of this approach:
- Update claude-code anytime — extensions are outside the CLI package
- Skills are just markdown — no build step, no breaking changes
- Hooks intercept everything — pre/post tool use without CLI modification
- Works with cloud or local — Anthropic Pro for power, Ollama for offline/free
- Playground plugin — interactive HTML prototyping built in
- Phase 1 — Git repo, launcher, CLAUDE.md, gitignore
- Phase 2 — Data layer: PromptStore, EmbeddingStore, DatasetGenerator, WebSearch, ArXiv, Classifier
- Phase 2 — Model manager: OllamaAPI, ModelfileBuilder, GGUFManager, LoRAManager, HFDownload
- Phase 2 — CoreCoder prompt dataset (19 rows) +
corecoder:latestmodel - Phase 2 — 11 slash commands + 6 custom skills
- Phase 2 — Speech pipeline (Whisper STT → LLM → edge-tts)
- Phase 2 — Code graph server (FastAPI + Cytoscape.js)
- Phase 2 — Agent-chef dataset generation pipeline
- Phase 2 — Playground plugin integration
- Phase 3 — Custom plugins (commands + agents for project archetypes)
- Phase 4 — Hook system (automated code review, safety checks)
- Phase 5 — MCP server integration
Built on top of Claude Code by Anthropic. Skills reference from the official skills repo. All customization in this repo is original work.