Claudemd Architect — AI skill for Claude Code

claudemd-architect

Generate production-grade CLAUDE.md files that steer AI coding agents. Incorporates LLM reasoning failure mitigations to prevent architectural drift, scope creep, and silent bugs in Claude Code sessions.

Based on [Large Language Model Reasoning Failures](https://arxiv.org/abs/2602.06176) (Song et al., TMLR 2026).

---

The Problem

A CLAUDE.md is not just documentation — it's a **steering mechanism** for an AI coding agent. A poorly written one leads to:

• 🔀 Architectural drift — the AI gradually deviates from your intended design
• 🐛 Silent bugs — errors that compile and run but produce incorrect results
• 📦 Scope creep — unauthorized frameworks, abstractions, and "helpful" additions
• 🔄 Wasted iterations — fixing AI mistakes that a better prompt would have prevented

The Solution

This skill produces CLAUDE.md files designed around **five LLM failure prevention principles:**

Principle	What It Does	Which Failure It Prevents
Constraints over suggestions	MUST / MUST NOT rules, not recommendations	Confirmation bias, proactive initiative
Rationale with every decision	ADRs with "why" and rejected alternatives	Reasoning drift, inverse inference
Concrete code examples	Exact patterns, not abstract descriptions	Compositional reasoning failures
Explicit anti-patterns	What NOT to do, mapped to failure categories	All failure modes
Verification checkpoints	Stop-and-verify gates at phase boundaries	Working memory limitations

---

Installation

Claude Code (global — recommended)

git clone https://github.com/keskinonur/claudemd-architect.git
cp -r claudemd-architect ~/.claude/skills/claudemd-architect

Now available in all your Claude Code projects as `/claudemd-architect` or triggered automatically when you mention creating a CLAUDE.md.

Claude Code (project-specific)

cp -r claudemd-architect /path/to/your/proje