TL;DR: guide-LLiMes is a framework for writing LLM coding guidelines that models actually follow. Instead of prompting tricks, it uses structured documentation with tiered rules (Must/Should/Nice) that you expand based on real failures.
Weâre excited to introduce guide-LLiMes, a framework for building LLM coding guidelines that actually work.
The Problem
If youâve worked with LLMs on code, youâve probably experienced:
- Inconsistent output quality
- Hallucinated dependencies or patterns
- Code that âworksâ but violates project conventions
- Instructions that get ignored or forgotten
Sound familiar? Youâre not alone.
Our Approach
guide-LLiMes is built on a few core principles:
Context Over Tricks
Rather than relying on clever prompting hacks, we focus on giving models stable project facts: scope, stack, structure, and constraints. Models perform better when they understand the terrain.
Tiered Guidance
Not all rules are equal. We use a Must / Should / Nice tier system:
- Must: Non-negotiable rules that prevent breakage
- Should: Guidelines that improve consistency
- Nice: Maturity practices for scale
This lets you start minimal and expand based on real failures, not hypothetical ones.
Tool-Agnostic Core
The principles work across Claude Code, Cursor, Copilot, and other tools. Only the file format and loading behavior change.
Getting Started
-
Assess your current state â Use the framework checklist to see what documentation you already have and whatâs missing.
-
Understand the methodology â Browse the fundamentals to learn how structured guidance differs from ad-hoc instructions.
-
Run the audit skill â If youâre using Claude Code, the
llimes-auditskill can analyze your existing documentation and identify gaps.
Whatâs Next
Weâre building out:
- Skills for auditing and improving your documentation
- Tool guides for specific IDEs and workflows
- Patterns drawn from real-world usage
This is an open project. Contributions and feedback welcome.