Top AI Repos — open-source AI, indexed and scored
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
Unofficial playbook for production-grade LLM system prompt architecture, derived from local analysis of Claude Code prompt patterns.
| Date | Stars |
|---|---|
| 2026-07-31 | 278 |
| 2026-08-06 | 278 |
Today
— stars today
This week
— stars this week
This month
— stars this month
Momentum
0.0
growth rate 0.00%/day
# Unofficial Claude Code Prompt Playbook > An unofficial, local-source-derived playbook for building production-grade system prompts, tool prompts, memory prompts, verifier prompts, and prompt architectures for modern LLM applications. ## Podcast Companion This repo also includes a podcast companion episode: - [`How_Anthropic_engineers_Claude_system_prompts.m4a`](./How_Anthropic_engineers_Claude_system_prompts.m4a) The audio is meant to accompany the playbook and give a higher-level walkthrough of why these prompt-architecture patterns matter. ## What This Playbook Gives You We use the learnings from the analyzed prompt architecture to build a practical, production-grade manual for creating world-class prompts and agent systems. This playbook is designed to help you build: 1. A gold-standard system prompt architecture for modern LLM applications 2. A modular way to break prompts into reusable policy blocks instead of one giant blob 3. A clean separation between static prompt policy and dynamic runtime context 4. A safe model for injecting user data only when it has causal value 5. A clear instruction-precedence system for base rules, org rules, project rules, user preferences, and request-time overrides 6. A strong trust-boundary model for policy, trusted runtime context, retrieved evidence, and user input 7. Better tool prompts, tool schemas, and tool-use policies 8. Verifier-agent patterns for adversarial, evidence-based validation 9. Durable memory patterns that stay useful instead of becoming noisy 10. Prompt-caching-friendly layouts that preserve a reusable prefix 11. Example-driven prompt design for formatting, edge cases, and escalation behavior 12. Anti-rationalization rules that explicitly block the model's most common shortcuts 13. Real-world templates for coding agents, RAG analysts, support agents, and orchestrators Think of the rest of the document as a chapter-by-chapter build manual for those outcomes. ## Disclaimer This document is based primarily on static analysis of an unofficial extraction of Anthropic's Claude Code prompt architecture. It is **not** an official Anthropic document and should not be read as a claim about hidden or proprietary policy. The point of this playbook is practical: - extract the strongest prompt-engineering patterns visible in a real production agent - explain why those patterns matter - generalize them into reusable templates for new LLM applications Where public Anthropic or OpenAI docs agree with the local observations, those are cited as secondary corroboration. ## Why This Exists Most prompt advice still lives at the level of: - "be clear" - "give the model a role" - "add examples" - "be specific" That advice is correct, but incomplete. A real production prompt stack is not a paragraph. It is closer to an operating system: - a top-level policy layer - runtime context injection - tool-specific instructions - specialist sub-prompts - memory and durable instructions - caching boundaries - verification gates - explicit trust boundaries The prompt system analyzed here strongly suggests that the best prompts are engineered like infrastructure, not written like marketing copy. ## The Most Important Observation The biggest lesson is this: > The world's best system prompts are not single prompts. They are layered instruction architectures. The analyzed prompt stack behaved as if it were made of: - a static policy core - a dynamic runtime tail - durable user and project instructions - tool-specific operating procedures - specialist prompts for planning, execution, summarization, and verification - evidence-handling rules - anti-rationalization rules That is the mental model this playbook uses throughout. ## What Was Observed In The Local Prompt Architecture ### 1. The system prompt is a policy stack, not a persona The observed prompt architecture was composed from multiple sections rather than one monolithic instruction block. The sections behaved roughly li
Excerpt of 37,865 characters
Read on GitHubWould you bet a product on this? Bounded 0–100 and slow moving.
matched fp:32f90953a9972453, llm:description: 'Unofficial playbook for production-grade LLM system prompt architecture, derived from local analysis of Claude Code prompt patterns.'