What it does This project documents 30 reconstructed prompt patterns and architectural behaviors from Claude Code, assembled purely through behavioral observation, output analysis, and community discussion. Think of it as a field guide to how a modern agentic coding assistant likely structures its inner monologue.

The interesting bit The authors treat the black box as a puzzle: they infer a modular prompt assembly pipeline with cacheable prefixes and dynamic per-session suffixes, plus a multi-stage security classifier for auto-approving tool calls. It’s forensic prompt engineering—useful precisely because it’s unofficial and unvarnished.

Key highlights

• Reconstructed patterns span core identity, orchestration, specialized agents, security classification, context compaction, and memory hierarchies
• Documents inferred sub-agents: verification (adversarial testing), explore (read-only), and even an “agent creation architect” that spawns new agents from requirements
• Proposes a memory loading order with transitive file inclusion and path-based conditional injection
• Explicitly disclaims being a leak; all content is approximate and invites correction

Caveats

• Everything is reconstructed approximation, not verbatim source; the README repeatedly warns that “actual implementation may differ significantly”
• No code to run—this is documentation and analysis, not a toolkit

Verdict Worth a skim if you’re building or hardening agentic tools and want a plausible reference architecture to argue with. Skip it if you need copy-paste prompts or verified internals.