Prompt Patterns: 10 Practices for Reliable Output

These are ten practices I rely on for getting reliable output from an AI system, with a short example for each. They build on the basics covered in Prompting Fundamentals.

#	Practice	What to do	Quick example
1	Pick the right "mode."	A conversational prompt is ad hoc and forgiving. A product or system prompt powers real software, so I treat it like production code: version it, test it, monitor it.	Product prompt for an email-draft feature: SYSTEM: Draft a friendly reply using {{tone}}. Include {{cta}}.
2	Use few-shot prompting.	I show a handful of input-output pairs that the model should mimic. This consistently improves accuracy on classification and formatting tasks.	Prompt: Classify ICD-10 codes. EX1: "Broken wrist" → S62.101 EX2: "Type 2 diabetes" → E11.9 User text: "Migraine without aura"
3	Decompose tough tasks.	I ask the model to break the problem into sub-steps before it attempts to solve it.	"First list the needed API calls, then write the Python script."
4	Add self-criticism.	I have the model review and improve its own answer before I see it.	"Provide a draft, then critique it in three bullets or fewer, then give a final improved version."
5	Feed rich, relevant context.	I prepend documents, background, or prior messages, with the most important material first.	Put the client's last three exchanges above the request so the model has the relevant history.
6	Ensemble when the stakes are high.	I run several differently worded prompts and take the majority or highest-confidence answer.	Three prompts each propose a formula; the version at least two of them agree on is the one I use.
7	Use chain-of-thought sparingly.	I let the model reason step by step only when I need to see that reasoning. Otherwise I ask it to hide the intermediate steps.	"Show your step-by-step reasoning, but put only the answer after a line that says FINAL:."
8	Skip "role" and "threat" tricks for accuracy.	Telling the model it is a Nobel laureate changes its tone, not its correctness, and threatening it tends to hurt rather than help. I reserve role prompts for voice and style, not for accuracy.	Creative writing prompt: "Write in the style of a formal legal memo."
9	Harden against prompt injection.	I assume that any text a user or a document supplies could contain instructions aimed at the model. I use retrieval-time filters, output checks, and logging rather than trusting the input.	Sanitize a user-supplied file name before it is included in a prompt that triggers a system action.
10	Iterate like software.	I test variants, log the failures, and refactor for brevity. Small wording changes can materially change the result.	Track prompt versions the same way I track code, and roll back a version if it starts producing more errors.