An Effective Prompt Structure – Ask, Context, Expectations (ACE)
Alex Hutanu
Principal Engineering Manager at Microsoft Azure Core (Cloud+AI) | GenAI Engineering | Technology Leader | Agile Leader
Generative Artificial Intelligence (AI) models like GPT-5 or Claude 4.1 Sonnet/Opus appear magical when they work and frustrating when they don’t.
Prompt engineering — the art of shaping your instructions or request so a model does what you need — has emerged as an essential skill. A 2025 guide notes that most prompt failures come from ambiguity rather than model limitations, and OpenAI’s own documentation emphasizes that the way you phrase the instruction, context and output format strongly influences the result. Overly vague or poorly scoped prompts encourage hallucinations—confident but incorrect responses.
Recent research on hallucination mitigation shows that providing clear instructions or specific context reduces hallucinations and that structured output generation enforces logical consistency.
And here the simple Ask, Context, Expectations (ACE) structure comes in to help - To set your Generative AI (GenAI) model for success, you essentially need to tell it what to do (what you Ask for), give the necessary background information (the Context details that influence this) and specify what a good deliverable looks like (what do you Expect to get back).
While advanced techniques exist, I’d like to show you a powerful yet simple structure that can improve results, by organising your prompt into those exact three sections:
- Ask – the task and purpose.
- Context – all relevant details and background information.
- Expectations – the desired output format.
Ask – Define the task
The Ask is your directive. It is the instruction that tells the AI exactly what task it should perform. Without a clear directive, the model may produce generic or irrelevant answers.
OpenAI advises placing the instruction at the beginning of the prompt and separating it from context using markers.
Being specific—“summarize the following customer support chat in three bullet points focusing on the issue, customer sentiment and resolution” rather than “write a summary”—dramatically improves the output.
Research on hallucinations underscores this point: explicit, context-rich instructions reduce the model’s creative freedom and limit hallucination.
When crafting the Ask:
State the purpose and scope. Describe what you want the AI to accomplish and why. Ambiguous tasks invite the model to guess.
Use action verbs and avoid vagueness. “Generate,” “list,” “explain” or “compare” signal exactly what you expect. Avoid imprecise phrases such as “talk about” or “fairly short”.
Specify roles if needed. Certain models perform better when assigning a role guides tone and content. For example, telling the model “You are a physics professor” produces a more accurate explanation.
A precise Ask frames the problem and prevents the model from wandering off into unrelated topics.
Context – Provide relevant background
The Context gives the model the information it needs to succeed. Lakera’s 2025 report reiterates that clear structure and context matter more than clever wording. Research on hallucinations confirms that augmenting system prompts with contextual metadata helps the model avoid misinterpretations, and that structured context reduces erroneous outputs.
To build effective context:
- Include input data and relevant facts.
If the prompt relies on a document, question or dataset, paste it here. Some models cannot access external information unless you provide it.
- Avoid overload.
Irrelevant details dilute the signal. Including only necessary facts reduces token waste and helps the model focus. More relevant for programatic usage (e.g., developing AI agents).
Think of context as the briefing materials you would give a human colleague: enough information to understand the problem but not so much that they get lost. Ask yourself if it’s relevant in any way to producing your expected results, and remove if not. Although, with today’s large context windows of GenAI models (ranging from 200k on frontier models to even millions), in non-programatic usage, I’d say your focus should be to not omit any relevant context rather than cutting out.
Expectations – Specify the deliverable
The Expectations section tells the AI how to present its answer. Output formatting ensures the response follows a particular structure—whether it’s a list, a table or a paragraph. Without clear formatting instructions, the AI may produce technically correct but unusable output. OpenAI similarly recommends articulating the desired output format through examples and being specific about length, style and other parameters. Azure’s prompt-engineering guidance highlights priming the output—adding cues such as “Here’s a bulleted list of key points:” to ensure bullet-pointed responses—and using clear syntax with section markers like — to separate information. In the hallucination research, structured output generation—requiring the model to produce code or structured data before a natural-language answer—reduced hallucinations by enforcing logical consistency.
When setting expectations:
- Choose an output format.
Specify whether you want a bullet list, numbered steps, a JSON object, a table or a narrative paragraph. If you need machine-readable output, ask for CSV or JSON (some GenAI inference platforms, such as Azure AI Foundry, support even structured outputs – which limit the model’s response to a specific format).
- Provide a template.
Show the AI the desired structure with placeholders. This is especially effective for structured formats.
- Add examples.
Few-shot examples show the model the structure you expect and improve accuracy. Use them always when the task is complex or the desired format is unusual. Consistency it’s also improved with example diversity. General rule - the more, the better.
- Define tone and style.
State if the output should be formal, conversational, concise or explanatory. GenAI models can adapt to specific audiences when guided.
- Set length and scope limits.
Indicate word counts or paragraph limits to prevent overly long responses. Although rarely working with precision (as GenAI models work with tokens not words), it achieves a better outcome than not provided.
Recommended by LinkedIn
- Clarify citation or source requirements.
If the model is connected to a retrieval system or grounded in data or provided with web search capabilities/tools, ask it to cite sources; if not, caution that requesting citations may cause hallucinations.
- Use section markers or metadata tags.
Delimit input sections and output cues with ###, — or XML/Markdown tags to improve parsing.
By spelling out your expectations, you eliminate ambiguity and make post-processing easier. In data analytics, for example, researchers found that forcing the model to produce structured output before natural language answers reduced hallucinations and improved accuracy.
Bringing it all together
Putting these sections into practice is straightforward.
Below is a generic template you can adapt:
Ask
You are a [role/persona]. Your task is to [primary action] for [purpose/audience].
Context
[Provide necessary background information, data, and any constraints. Include all that is relevant.]
Expectations
- Output format: [e.g., bullet list, JSON, table, paragraph].
- Tone/style: [e.g., formal, accessible, playful].
- Length: [word or paragraph limit].
- Examples: [the more diverse, the better]
- Any other specific requirements. (e.g., Include/omit citations or sources)
This structure does not replace all (and can be used alongside) prompt-engineering techniques but it provides a dependable foundation. Good prompt engineering can dramatically improve output quality without retraining or adding more data, and prompts shape not just content but tone, structure and safety. Treating the AI like a new hire—giving it a clear task, relevant context and precise deliverable—sets it up for success.
Final thoughts
The Ask, Context, Expectations (ACE) structure is both simple and powerful. It distills the essence of prompt-engineering into a manager’s checklist: what do I need done, what information do you need, and what does success look like?
In my experience so far, following this structure reduces ambiguity, mitigates hallucinations and yields more consistent outputs. Whether you are automating reports, summarising conversations or brainstorming ideas, investing a few extra minutes in crafting a structured prompt pays dividends.
I also found myself using it as a mindset in my personal interactions with GenAI as well – when writing Copilot to start a “deep research” for example, I’m questioning if my Ask is clear enough, and if I’ve provided enough Context to complete the Expected results)
For those interested in digging deeper, explore advanced techniques such as chain-of-thought prompting, reason and act, or role prompting. But if you just want better answers from AI models without becoming a prompt engineer, remember: Ask clearly, provide Context and set Expectations.
Appendix
Prompt engineering techniques - Azure OpenAI | Microsoft Learn
Best practices for prompt engineering with the OpenAI API | OpenAI Help Center
Beyond Fine-Tuning: Effective Strategies for Mitigating Hallucinations in Large Language Models for Data Analytics
The Ultimate Guide to Prompt Engineering in 2025 | Lakera – Protecting AI teams that disrupt the world.
Understanding Prompt Structure: Key Parts of a Prompt
What is Hallucination, and How Can It Be Controlled Using Prompt Engineering?
Optimistic Tech Leader, Empowering Teams through Innovation and AI
2moThat is a great strategy, Alexandru Hutanu ! Thank you for sharing. What I do in addition to proper structuring the prompt is treating it as a conversation. I follow-up with feedback, clarification, ask it to evaluate and rate its previous response. While I learned this before reasoning models came along, which already improve this process a lot, I still find this gives results closer to my needs. However perfect the prompt might be, since this isn't a deterministic approach, "you never know what you're gonna get" 🙂