Should-Have Guidelines

These guidelines reduce variance. The LLM can produce working code without them, but the output will be less consistent, harder to review, and more likely to drift from your team’s established patterns.

Add these after your must-haves are solid — typically within the first month of using LLM-assisted coding on a project.

The Seven Should-Haves

#	Guideline	What it improves
1	Spec Templates	Consistent prompts, nothing gets forgotten
2	Architecture Decision Records	Model understands why things are the way they are
3	Review Guidelines	Faster reviews, clear expectations
4	Error Handling	Uniform error taxonomy and propagation
5	Performance Expectations	Model considers scale, avoids naive implementations
6	Test Data Conventions	Reproducible, isolated tests
7	Documentation Structure	Docs go in the right place, stay findable

Prioritizing

Pick based on where you see the most friction:

• LLM output varies wildly per request → Spec Templates
• Model keeps suggesting patterns you rejected → ADRs
• Reviews take too long → Review Guidelines
• Error handling is inconsistent → Error Handling
• Tests are flaky or interdependent → Test Data Conventions

References

• External References

---

Table of contents

• Spec Templates
• Architecture Decision Records
• Review Guidelines
• Error Handling
• Performance Expectations
• Test Data Conventions
• Documentation Structure