Three-Tier Model

Not all guidelines carry equal weight. Some prevent real breakage. Others improve consistency. A few only matter at scale. Treating them all as equally urgent leads to either paralysis (too much to write) or a bloated guidelines file that the model can’t reliably follow.

guide-LLiMes organizes guidelines into three tiers based on a single question: “If we skip this, what breaks?”

The Tiers

Must-Have — Prevent breakage

These are non-negotiable. Without them, the LLM will produce code that fundamentally doesn’t fit your project.

Guideline What breaks without it
Project Scope Model invents features outside your domain
Tech Stack Model suggests wrong languages, frameworks, libraries
Coding Standards Every generation looks different, diffs are noisy
Testing Strategy Model skips tests or writes wrong types
Directory Structure Files land in wrong places, imports break
API Contracts Model breaks public interfaces, ignores versioning
Security Basics Model logs secrets, bypasses auth
LLM Guardrails Model modifies migrations, CI configs, infra

When to write: Stage 1. Pick at least 3 that matter most for your project.

Should-Have — Reduce variance

These make LLM output more predictable and reviewable. The model can produce working code without them, but the quality and consistency will fluctuate.

Guideline What improves with it
Spec Templates Prompts become consistent, nothing gets forgotten
ADRs Model understands why things are the way they are
Review Guidelines Reviews are faster, expectations are clear
Error Handling Errors are typed, logged, and handled uniformly
Performance Expectations Model considers scale, doesn’t write naive implementations
Test Data Conventions Tests are isolated, reproducible, not flaky
Documentation Structure Docs go in the right place, stay findable

When to write: Within the first month. Prioritize based on where you see the most variance in LLM output.

Nice-to-Have — Scale and compliance

These matter when your team grows, your domain is regulated, or you want to prevent long-term architecture drift.

Guideline When it becomes necessary
CI Quality Gates Team > 3 people, or many LLM-generated PRs
Architecture Tests Module boundaries are being quietly violated
Domain Glossary Model invents conflicting terms for the same concept
Migration Rules Schema changes are frequent and risky
Security Playbook Domain handles sensitive data or is regulated
LLM Evaluation Loop You want to systematically improve guideline effectiveness

When to write: First quarter onward, driven by actual pain points.

How to Use This Model

  1. Start with must-haves. Pick the 3 that address your biggest pain points. Write them. Use them for a week.
  2. Observe what goes wrong. If the LLM keeps making the same type of mistake, there’s a missing guideline for it.
  3. Promote as needed. Something you thought was “nice-to-have” might turn out to be a must-have for your project. The tiers are starting points, not fixed categories.
  4. Don’t write everything at once. A guidelines file with 50 rules that you wrote in one sitting is almost certainly too broad, too shallow, or both.

Tier Boundaries Are Project-Specific

A healthcare startup handling patient data will need Security Basics as a deep, detailed must-have from day one. A weekend side project might skip it entirely.

A team building a public API will treat API Contracts as critical. An internal tool with no external consumers might keep it minimal.

Use the tiers as a framework for thinking about priority, not as universal truth.


Next: Token Budgets — why your guidelines can’t be infinitely long.

References


Back to top

guide-LLiMes — Build LLM coding guidelines that work.

This site uses Just the Docs, a documentation theme for Jekyll.