The Prompting Pattern That Stuck After A Year

I have tried every prompting technique that made the rounds. Chain of thought. Tree of thought. Self consistency. Role prompts. Most of them worked for a week and then faded as models got better.

The business problem is reliability. When a client is on the line, prompts need to hold up. You cannot ship a feature that only works on Tuesdays.

The pattern

State the artifact. Then state the constraints. Then state the context. In that order.

Artifact: A Python function that validates a UK VAT number.
 
Constraints:
- Pure function, no network calls
- Returns (is_valid: bool, reason: str)
- Handles both GB and XI prefixes
- No external libraries
 
Context:
This runs inside a Django form validator. It must be safe for untrusted input.

That is it. I call it ACC. Artifact, constraints, context.

Why it survives

Artifact forces concreteness. The model cannot wander off because the shape of the output is pinned. Constraints eliminate the most common failure modes before they happen. Context handles the last mile of domain specificity.

What I dropped

• Personas like 'You are a senior Python engineer'. Models no longer need the hint.
• Step by step instructions. The better models plan on their own.
• Few shot examples for common tasks. They bloat the prompt and rarely help now.

A small story

I rewrote a client's prompt library to ACC shape last quarter. Their success rate on a business validation agent went from 78 to 94 percent. No model change. Same budget.

Closing

See the Anthropic prompt engineering guide for the canonical version. My ACC pattern is a compressed flavor of it. It fits on a sticky note. That is why it stuck.