Claude Code System Prompts Research
Behavioral reconstruction of agentic coding assistant prompt architecture, coordination, and security — a blueprint for building your own.
- Source repo
- Leonxlnx/agentic-ai-prompt-research
- Stars
- ★ 2.6k
- Last updated
- 6mo ago
- FA score
- 35/100 · Major gaps
At a glance
- How it runs
- Works with
- Platform-specificClaude Code
- Setup effort
- Low · running in minutes
- You'll need
- Typical use
- AI engineers building their own agentic coding tools can reference the dynamic prompt assembly and context management patterns.
- Not a fit if
- Teams needing runnable code or integration APIs
- Teams needing authoritative verbatim official prompts
- Teams needing a clear open-source license before reuse
- Source review
- 35/100 · Major gaps 2 safety controls not found
What does this agent do, and when should you use it?
This repository is an independent research project that reconstructs the prompt architecture, agent coordination, and security mechanisms of modern agentic coding assistants like Claude Code, based on behavioral observation, output analysis, and community discussions. It documents approximately 30 patterns across core identity, orchestration, specialized agents (verification, explore, agent creation), security classification, tool descriptions, context window management, dynamic behaviors (proactive mode, browser automation), and skill patterns (simplify, skillify, stuck, remember). The project emphasizes that all content is reconstructed approximation, not verbatim copies, and is not affiliated with Anthropic. Its value lies in helping AI engineers, prompt engineers, and security researchers learn from production-grade agent design patterns and apply them to their own projects.
The repository provides a series of Markdown documents, each detailing a prompt pattern. For example, prompts/01_main_system_prompt.md explains how the master prompt is dynamically assembled via a pipeline of modular builders with a cache boundary, and prompts/12_yolo_auto_mode_classifier.md describes a multi-stage security classifier for autonomous tool execution. The documentation covers how to build modular system prompts, coordinate multiple sub-agents, define base instructions (default agent prompt), manage tool permissions via risk assessment, compact conversation context, and load hierarchical memory (enterprise config, user global prefs, project-level, local overrides). It contains no executable code; it is purely educational reference material.
- AI engineers building their own agentic coding tools can reference the dynamic prompt assembly and context management patterns.
- Prompt engineers studying production-grade architectures can learn from modular builders, cache boundaries, and security classifiers.
- Security researchers analyzing how autonomous AI tools manage permissions can examine the auto-mode classifier and permission explainer patterns.
- Students and educators teaching multi-agent system design can use the documented orchestration and specialized agent patterns.
- Developers wanting to implement memory hierarchies or permission overrides can adopt the memory selection and permission explainer designs.
How do you install or deploy this agent?
No installation is required as this is a research repository without executable code. Users can clone the repository (e.g., git clone https://github.com/Leonxlnx/agentic-ai-prompt-research) or browse the prompts/ directory directly on GitHub. No dependencies or configuration needed.
How do you use this agent?
Start by reading the README.md catalog, then navigate through prompts/ in numeric order. For instance, 01_main_system_prompt.md covers main prompt assembly, 05_coordinator_system_prompt.md covers orchestration, and 12_yolo_auto_mode_classifier.md covers security classification. Since this is documentation, integration requires extracting patterns and implementing them in your own project; there is no API or CLI.
What are this agent's strengths and limitations?
- Comprehensive catalog of ~30 patterns covering full agent lifecycle from identity to skill management.
- Includes architectural observations (e.g., prompt assembly pipeline, cache boundary, memory hierarchy) that deepen understanding.
- Explicitly labeled as reconstructed approximations, reducing copyright concerns.
- Concrete documentation facilitates direct translation to engineering practice.
- Reconstructions are speculative and may differ significantly from actual implementations; not authoritative.
- No official endorsement, accuracy not guaranteed.
- No executable code or integration interfaces; requires self-implementation.
- Lacks an open-source license, so reuse risk must be assessed.
How does this agent compare with similar options?
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Claude Code System Prompts Research This agent | 35 · Major gaps | Library / SDK | ★ 2.6k | 6mo ago | — | Claude Code |
| PowerContext | 55 · Major gaps | CLIFree | ★ 1.3k | today | Python | Codex · Claude Code |
| 12-Factor Agents | 24 · Major gaps | Library / SDKFree | ★ 27k | 1y ago | TypeScript | — |
| Loop Engineering | 69 · Some gaps | CLIFree + model costs | ★ 12k | today | TypeScript | Codex · Claude Code |
How does FollowAgents rate this agent?
Why each dimension lost points
Evidence shows discussion of permission classification, user confirmation, data flow transparency, sensitive data handling, and external effects, but mostly descriptive without implementation details. Dependency security and rollback mechanisms are not mentioned. Source attribution is stated but unverified.
Internal consistency is good, documentation structure is clear, but no dependency management or failure handling is described.
Target audience and use cases are clear, capability boundaries are described, but trigger precision and environment fit are insufficient.
Information architecture is clear, naming is stable, but installation notes, license, versioning, and maintenance responsibility are missing.
Output usability is high, marginal value is clear, but cost-benefit analysis is lacking.
Claims are based on observation and inference, but lack traceable sources; fact-inference separation is good.
- Not found in source: dependency securityPin versions and run a dependency audit (npm audit, pip-audit) before installing; prefer running it in a container.
- Not found in source: rollback or recovery pathBack up first, or work on a git branch or snapshot, so its changes can be undone.
- This is a research project with reconstructed approximations, not official documentation; use with caution.
- License information is missing, which may affect compliance.
- No installation or integration guide is provided; practical application requires self-exploration.