Prompt Engineering Patterns That Actually Work

There is a lot of prompt engineering content online that reads like it was written by someone who has experimented with a chatbot for an afternoon. This piece is different: these are patterns I have validated across production applications handling millions of requests.

Some of what is widely cited does not hold up under real production conditions. Some things that do not make intuitive sense actually work. Here is what I know to be true.

Pattern 1: Role Assignment With Expertise, Not Job Title

The common advice: "You are a helpful assistant." The better version: "You are a senior data analyst with 10 years of experience in financial services, reviewing reports for institutional investors who have graduate-level quantitative backgrounds."

The specificity does two things. It contextualizes the appropriate vocabulary and depth. And it sets a quality expectation - an expert reviewer produces different output than a generic assistant.

But the pattern only works when the role is genuinely relevant. Assigning a role that has nothing to do with the task adds tokens without adding value.

Pattern 2: Specify the Anti-Pattern

Instead of only telling the model what to do, tell it what NOT to do. This is underused and highly effective.

"Summarize this article. Do not include opinions about the author's argument. Do not use bullet points. Do not exceed 150 words."

The negative constraints are often more reliable than positive ones because they bound the space of acceptable outputs more precisely. I use this pattern in every production prompt where output format consistency matters.

Pattern 3: Chain of Thought for Complex Reasoning

For tasks that require multi-step reasoning, explicitly instruct the model to work through the problem before giving a final answer: "Think through this step by step before providing your final response."

The improvement in accuracy on reasoning tasks is well-documented in research and consistently reproducible in practice. The mechanism is that the model uses its token generation to reason before concluding, rather than jumping to a conclusion.

Practical note: this increases token usage by 30-50% on average, which affects cost. For tasks where reasoning quality matters, it is almost always worth it.

Pattern 4: Output Format as a Template

Rather than describing the desired output format in prose, provide an empty template:

"Respond in this exact format:\n\nSUMMARY: [2-3 sentences]\n\nKEY POINTS:\n- [point 1]\n- [point 2]\n- [point 3]\n\nCONFIDENCE: [High / Medium / Low]"

Format templates dramatically reduce output variance compared to prose format descriptions. When you are building applications with ChatGPT or Claude where you need to parse the response programmatically, format templates are essential.

Pattern 5: Few-Shot Examples for Tone and Style

For creative or style-sensitive tasks, examples outperform instructions. If you want a specific writing voice, provide 2-3 examples of that voice and ask for output "in the same style."

The model is much better at pattern-matching to concrete examples than at interpreting abstract style instructions like "write in a professional but approachable tone." What you think "approachable" means and what the model generates for "approachable" can be quite different.

Limitation: examples consume context tokens. For long-form content where context space is limited, you may need to abbreviate or select examples carefully.

Pattern 6: Iterative Refinement in the Prompt

For complex outputs, build the refinement loop into the prompt: "Generate a first draft of X. Then review it and identify 3 specific weaknesses. Then produce a revised version that addresses those weaknesses."

This self-critique approach significantly improves output quality for many tasks. The model generates, evaluates, and revises within a single inference - which is faster and often more effective than running multiple separate prompts.

I use this pattern for ChatGPT and Claude tasks involving writing, code review, and structured analysis where initial outputs tend to miss nuance.

Pattern 7: Explicit Uncertainty Handling

For factual tasks, add: "If you are uncertain about any claim, state your uncertainty explicitly rather than presenting it as fact."

This does not eliminate hallucination but it changes the failure mode: instead of confidently wrong assertions, you get marked uncertainties that a human reviewer can check. For production applications, this is a much better failure mode.

Combine this with a source citation requirement: "Cite the specific source or knowledge that supports each factual claim." The model cannot always do this accurately, but when it can, it gives you a verifiable reference.

What Does Not Work Well

Magic phrases: "Answer in the style of an expert" with no further specification is essentially meaningless. Specificity is everything.

Extremely long system prompts: There is a real degradation in instruction-following for very long system prompts (over 2,000 tokens). The model attends less reliably to early instructions. Keep system prompts focused.

Asking for multiple unrelated things: Prompts that combine a creative task with a factual task with a formatting task tend to under-deliver on all three. When possible, separate concerns into separate prompts.

The Meta-Pattern

Every effective prompt has three things: clear context (who is speaking, what is the situation), clear task (what do I need), and clear output definition (what does good look like). Everything else is optimization within that frame.

For more on choosing the right AI tools for different tasks, see best AI tools and the ChatGPT vs Claude comparison.