Skip to content

Prompt Design Architecture Decision Records (ADRs)

Architecture Decision Records are a lightweight method for capturing important design decisions along with their context and consequences. In software engineering, ADRs document choices about system architecture. Here we adapt the format for prompt engineering decisions -- the choices we make when designing, structuring, and orchestrating prompts in LLM-powered systems.

Why prompt design ADRs?

Prompt engineering involves more trade-offs than most teams realise. Should you use few-shot examples or fine-tune? One monolithic prompt or a multi-step pipeline? Inline safety checks or a separate validation pass? The answers depend on constraints that shift over time -- model capabilities, latency budgets, data sensitivity, and team expertise.

ADRs capture the reasoning behind a decision, not just the decision itself. Six months from now, when someone asks "why do we run two prompts instead of one?", the ADR explains the context that led to that choice and the alternatives that were weighed.

Template structure

Every ADR in this directory follows a consistent template:

Section Purpose
Status Accepted, Superseded, or Deprecated
Date When the decision was made
Context module Which curriculum module the decision relates to
Context The situation, problem, and constraints
Decision The specific approach chosen
Rationale Why this option over the alternatives
Alternatives Considered Other options, their pros/cons, why they were rejected
Consequences Positive outcomes, negative trade-offs, and risks
Related Decisions Links to other ADRs, modules, or prompt files

How to use these records

  • Learning: Read through the ADRs below to see how experienced prompt engineers reason about design trade-offs. Each record maps back to one or more curriculum modules so you can study the underlying principles.
  • Writing: When you face a non-trivial prompt design choice on a real project, copy the template and fill it in. The discipline of writing down alternatives and consequences often clarifies your thinking before you commit to an approach.
  • Reviewing: During prompt code review, reference relevant ADRs to explain why a particular pattern was chosen.

Table of contents

ADR Title Context Module
001 Few-shot prompting over fine-tuning for classification Core Principles / Patterns
002 Split monolithic prompt into planner + executor pipeline Advanced Patterns / Agentic Patterns
003 Add a separate safety-gate validation prompt Best Practices
004 JSON schema constraints over natural language format instructions Core Principles / Best Practices

Back to curriculum