Educational deep-dive into agentic coding assistants, documenting prompt patterns, orchestration, and security classifications. Based on public observations and reconstructions.
Collecting history — the radar snapshots this repo daily. The trend line appears after 3 days of data (1 so far).
What it is
Claude Code System Prompts is a research project that analyzes how modern agentic AI coding assistants operate. It documents prompt architecture, agent coordination patterns, and security mechanisms, focusing on how dynamic system prompts are built, how multiple sub-agents coordinate, and how tool calls are classified and approved.
How it works
The project outlines patterns across categories such as Core Identity, Orchestration, Specialized Agents, Security and Permissions, Tool Descriptions, Utility Patterns, Context Window Management, Dynamic Behaviors, and Skill Patterns. It presents an architectural view with examples like:
- Dynamic system prompts assembled from modular sections
- Multi-worker orchestration with phased workflows
- Adversarial testing and read-only exploration agents
- Multi-stage auto-mode classifiers for tool execution
- Memory, session, and memory-selection strategies
It also provides an architectural observation section detailing prompt assembly pipelines, a multi-stage security classification approach, and a memory hierarchy for configuration loading and local overrides.
Getting started
The repository is a research project with documentation of patterns and does not appear to include runnable setup or installation instructions in the provided README excerpt. The README lists documented patterns and describes the intended educational purpose.
Recent releases
There are no releases listed in the provided data (RELEASES (latest 0): - none).
Traction
Repository has 2499 stars and 1068 forks.
Behind the repo
There is no information in the provided data about a company or startup behind the repo.
Caveats
License is listed as none. The README states the work is based on behavioral observation and reconstructions rather than proprietary content, and disclaims affiliation with any specific company.