Tried Piebald-AI/claude-code-system-prompts: A Practical Prompt Engineering Reference

Piebald-AI/claude-code-system-prompts is gaining attention quickly, with +13 GitHub stars today. The reason is simple: it collects the internal building blocks behind Claude Code in a format developers can inspect, compare, and learn from.

This is not another coding assistant. It is a prompt reference containing all parts of Claude Code’s system prompt, 27 built-in tool descriptions, and sub-agent prompts for Plan, Explore, and Task workflows. It also includes utility prompts covering CLAUDE.md, compaction, status lines, magic documentation, WebFetch, Bash commands, security reviews, and agent creation.

The repository is updated for each Claude Code version, making it especially useful for tracking how agent behavior evolves. I see it as a practical reference for building custom IDE agents, debugging unexpected model behavior, or designing safer tool-use workflows.

Quick Gateway Test

For an Anthropic-compatible relay, I would start with a minimal environment configuration:

export ANTHROPIC_BASE_URL="https://b-lost.com/v1"
export ANTHROPIC_API_KEY="your-b-lost-api-key"
export ANTHROPIC_MODEL="claude-fable-5"

For clients that use JSON configuration, the equivalent pattern is:

{
  "provider": "anthropic",
  "baseUrl": "https://b-lost.com/v1",
  "model": "claude-fable-5",
  "apiKey": "${B_LOST_API_KEY}"
}

B-Lost supports native Anthropic /v1/messages, so Prompt Caching can be used for repeated system prompts and tool definitions. With a 90% cache-hit discount, large agent instructions become much less expensive during iterative development. The relay also advertises 20% off official list pricing, which is useful when testing prompt-heavy IDE workflows.

My main takeaway: this repository turns Claude Code’s hidden complexity into something developers can study. If you are building a Cursor, Cline, Roo Code, or custom OpenAI-compatible workflow, it is an excellent prompt architecture reference—not just a collection of snippets.