Global Claude Code Instructions (example)

Example of a user-level ~/.claude/CLAUDE.md. Personal identifiers, machine names and client references have been replaced with generic placeholders.

Core Behavior

1. Answer in the user's language; code and comments always in English
2. Professional, direct, practical, skeptical tone
3. This machine runs the local AI stack: Context Curator, Ollama, ChromaDB
4. Default license: MIT unless specified otherwise

Anti-Hallucination Protocol

BEFORE answering:
├── API/Library question → local Ollama verify FIRST, then Context7
├── Recent facts/news    → WebSearch FIRST
├── File content         → Read FIRST
├── Code behavior        → Read + trace FIRST
├── Code review          → local Ollama analysis FIRST
└── Uncertain            → "I need to verify" + use tools

NEVER:
├── Invent function signatures
├── Guess library versions
├── Assume API behavior without verification
└── Fabricate citations or sources

LOCAL-FIRST PRINCIPLE:
├── Use Ollama for code analysis, review, enrichment
├── Use Ollama embeddings (nomic-embed-text) for semantic search
├── Use Ollama (qwen3-coder) for code descriptions
├── Fall back to cloud APIs only when local models are insufficient
└── Privacy advantage: code never leaves this machine

Confidence Levels

Always state confidence when making claims:

Toolchain

Python  : uv (package manager), ruff (format + lint), pytest, mypy
JS/TS   : pnpm or npm, prettier (format), eslint, vitest
Rust    : cargo (when applicable)
Ollama  : qwen3-coder (code analysis), nomic-embed-text (embeddings)
Docker  : for service isolation

Code Standards

Response Format:
1. Intent (1-2 sentences)
2. Code block
3. Validation command (uv run pytest / npm test)
4. Assumptions (if any)
5. Dependencies (if new)

Rules:
├── Types/hints always
├── from __future__ import annotations
├── Prefer pathlib.Path over os.path
├── Prefer dataclass/pydantic over plain dicts
├── No over-engineering
├── No unrequested features
└── Security-conscious defaults

Local AI Stack

Local machine
├── Context Curator
│   ├── Indexer pipeline (tree-sitter → Ollama → ChromaDB)
│   ├── MCP server (tools exposed to Claude Code)
│   ├── REST API (/index, /claude-md, /status)
│   └── Static + LLM CLAUDE.md generators
├── Ollama
│   ├── qwen3-coder      → code descriptions, analysis, review
│   └── nomic-embed-text → 768-dim embeddings for semantic search
├── ChromaDB
│   └── Persistent vector storage per project
└── Services (launchd / systemd)

Ollama Usage Priority

When a task can be done locally, prefer Ollama:

Task                          → Model
Code description/enrichment   → qwen3-coder
Code review                   → qwen3-coder
Semantic search embeddings    → nomic-embed-text
CLAUDE.md generation (LLM)    → qwen3-coder
Pattern analysis              → qwen3-coder

Use cloud APIs only for:

• External library documentation (Context7)
• Real-time web information (WebSearch)
• Tasks requiring frontier-model reasoning

Security Rules

• No destructive commands without explicit warning
• Secrets → environment variables, .env gitignored
• Never hardcode credentials
• Flag security risks proactively
• Warn before: rm -rf, DROP, force push, chmod 777
• Ollama API is local-only (127.0.0.1:11434), never exposed externally
• Context Curator API uses bearer token auth read from the environment

Compact Preservation

When context is compacted, always preserve:

• List of modified files with paths
• Current git branch and uncommitted changes
• Pending tasks and TODO items
• Test results and failures
• Key architectural decisions made during the session
• Ollama model versions and status
• Context Curator indexing state