claude code system prompts

The Problem

Software‑engineering AI assistants need a reproducible, auditable prompt stack to guarantee consistent behavior, security boundaries, and coordination among sub‑agents. Teams building similar assistants often reverse‑engineer these stacks from scattered examples, which is time‑consuming and error‑prone.

What This Does

The repository is a curated collection of Claude Code system prompts. Each markdown file (01_main_system_prompt.md … 30_prompt_suggestion.md) captures a distinct logical slice—core identity, orchestration, specialized agents, security classifiers, and utility helpers. The README.md indexes the catalog and explains how Claude Code assembles the prompts at runtime: a static “global prefix” plus a session‑specific “suffix” that can be cached across API calls.

Because the repo contains only static prompt text, there is no executable code, build scripts, or dependency files. Its purpose is to serve as a reference artifact for anyone who wants to reproduce Claude Code’s prompt architecture or adapt its patterns to another model.

How It Is Wired

There is no internal call graph to map: the repository does not include any source code that performs prompt assembly, API calls, or runtime logic. The only “wiring” present is the logical ordering documented in the README and the filenames themselves. For example:

If a consumer wishes to replicate Claude Code’s behavior, they must write their own orchestration code that reads these markdown files, concatenates the appropriate sections, and sends the resulting prompt string to the Claude API. The repository itself does not own any side‑effects such as file I/O, network traffic, or database updates.

How To Use It

1. Clone the repository

   git clone https://github.com/moses-y/claude-code-system-prompts
   cd claude-code-system-prompts

1. Select the prompts needed for your use case. For a full Claude Code session, concatenate in order:

   01_main_system_prompt.md
   02_simple_mode.md (optional)
   03_default_agent_prompt.md
   04_cyber_risk_instruction.md
   05_coordinator_system_prompt.md
   … (additional agent files as required)

1. Load the markdown content in your host language (e.g., Python pathlib.Path.read_text()), build the final prompt string, and pass it to the Claude API endpoint you are using.

No additional configuration files, environment variables, or build steps are provided; the repo is pure documentation.

Real‑World Use

A development team building an internal code‑assistant can import the prompt catalog, modify 04_cyber_risk_instruction.md to reflect corporate policy, and feed the assembled prompt to Anthropic’s Claude endpoint. The assistant will then inherit the same multi‑agent coordination and security checks described in the original Claude Code design.

from pathlib import Path
prompt_parts = [
    Path("prompts/01_main_system_prompt.md").read_text(),
    Path("prompts/05_coordinator_system_prompt.md").read_text(),
    # add or drop files as needed
]
final_prompt = "\n".join(prompt_parts)

response = claude_client.complete(prompt=final_prompt, max_tokens=1024)
print(response.text)

Code Health & Issues

• Medium – No test files detected – repository‑wide, tests directory absent.
• Medium – No CI/CD pipeline detected – no .github/, Makefile, or other automation configs.
• Medium – No LICENSE file – usage and redistribution rights undefined.
• Medium – No lockfile or dependency manifest – cannot infer package manager or reproducible environment.

Static analysis found no committed secrets and no Dockerfile. The repo consists exclusively of markdown files; therefore, typical software‑quality metrics (linting, coverage) do not apply.

The Bottom Line

The repo delivers a thorough, well‑organized snapshot of Claude Code’s prompt architecture, valuable for research or as a baseline for custom agent design. Its usefulness is limited to documentation; there is no runnable code, test suite, or licensing clarity, so consumers must supply their own execution framework and legal review. Ideal for engineers who need prompt‑level insight rather than a plug‑and‑play library.