3.2 Prompt Engineering Techniques

3.2 Prompt Engineering Techniques

Quick Answer: Microsoft's AB-731 skill "Understand techniques of prompt engineering" is a catalog leaders should be able to name and govern: instructions, context, examples (few-shot), role, constraints, output format, iteration, decomposition, and grounding references. Microsoft 365 Copilot habits transfer as goal, context, expectations, and sources. When those techniques cannot supply private or changing facts, or cannot lock a behavior, you move to retrieval-augmented generation (RAG), fine-tuning, or a different tool—you do not invent undocumented slash-commands.

You still are not writing code. You are choosing a method, funding a gallery, and knowing when to stop prompting and start a data, model, or process project.

A leader's map of techniques

Microsoft Foundry's Prompt engineering techniques article (current Learn documentation for Azure OpenAI in Microsoft Foundry Models) describes prompt components that map to how a manager should brief an assistant:

Few-shot is not "the model learned our company." Foundry is explicit: examples change this inference, typically as example user and assistant turns after a system message. If you need the behavior every time without pasting examples, that is a custom instruction, a gallery prompt, or—when that is still not enough—fine-tuning.

Cues (Foundry) prime the first words of the answer ("Key points:" then a bullet). Leaders can think of this as starting the template so the model continues in the right shape.

Recency. Foundry warns that models can show recency bias: the end of the prompt may weigh more, so repeating the critical instruction at the end is a documented experiment. Microsoft Support's Copilot prompting page (February 2026) similarly says order matters, later parts are often emphasized, and if Copilot must use specific files, put that information last.

Positive instructions. Copilot is geared toward action. Microsoft Support: tell it what to do, including if-then exceptions, rather than a stack of "don't"s. Microsoft's Security Copilot prompting tips document the same idea for that product.

Give an out. If the source does not contain the answer, instruct "say not found" instead of forcing a guess.

Few-shot in human resources, without code

A people-leader wants every coordinator to classify leave emails into the same four buckets. A gallery prompt can combine instruction, role, examples, constraints, and format:

• Instruction / role: You are an HR operations coordinator. Classify each leave request as Sick, Parental, Bereavement, or Other.
• Examples (few-shot): "I need Tuesday off for a dentist extraction" → Sick. "Bonding time after the birth of our child" → Parental.
• Constraints and out: If the email does not fit, use Other and quote the phrase that blocked a clean fit. Do not invent medical details.
• Output format: One line: Bucket | quoted phrase | confidence: high or low.

That is technique stacking. It is still not a substitute for an HR system of record, and a human still reviews edge cases.

Decomposition in operations

A single prompt that says "audit this 40-page vendor contract, extract risks, draft the board note, and propose a negotiation script" is how teams get fluent mush. Decomposition:

1. Extract only clauses that mention liability, data residency, and termination.
2. Label each as confirmed in the text or not found.
3. Draft a one-page board note from the labeled list only.

Each step is reviewable. That is prompt engineering as process design, which is the level AB-731 tests.

Microsoft 365 Copilot habits that transfer

Microsoft Learn's business-user module states four elements of an effective Copilot prompt: context (why you need it and who is involved), goal (what Copilot should do; generative models imitate, so examples help), source (where to look; with a Microsoft 365 Copilot license you can ground in work data such as email, meetings, and chats), and expectations (audience, tone, delivery). Support articles use the same four ideas with a slightly different order. Teach employees: start with a goal; add the rest when the first draft is not specific enough.

Microsoft Learn Write an effective prompt documents three reference symbols in Copilot Chat. Name only what that page names:

• / — reference a file or folder (Learn's example pattern: summarize key points from /Q4MarketingStrategy.pptx).
• # — reference a topic or meeting; Microsoft labels this Premium and notes that work content such as meetings and chats requires a Microsoft 365 Copilot license.
• @ — mention a person, for example when collaborating in a shared Copilot Page.

If a control is not on a current Microsoft Learn or Microsoft Support page you can open, teach the technique (name the file, name the meeting) without inventing a slash-command.

Iterate. After a response, Microsoft's training says to select sources when they are offered and to send a follow-up rather than accepting a miss. Support: start broader for ideas, then add details; for insights, summarize a file then ask narrower questions.

Personalization (enablement, not a secret prompt hack). Microsoft Learn Personalize Copilot with memory and custom instructions: Custom instructions are a standing brief (role, bullets, no jargon) under Copilot Chat settings → Personalization. Copilot Memory can save details (your organization's admin controls whether this is available). Temporary chat avoids memories and custom instructions for a neutral session. Leaders should decide what may be stored and publish that in policy.

Prompt Lab is the gallery: discover Microsoft-curated prompts, save with the bookmark icon, edit by re-running (Microsoft: you cannot edit a saved prompt in place; save a new one and delete the original to avoid duplicates), share, and consume org prompts your admin publishes. Access is the prompt pills or … below the input box. With a Microsoft 365 Copilot license, Schedule this prompt can run a work-grounded prompt automatically (Microsoft documents a maximum of 10 scheduled prompts; optional connected experiences must be enabled). Scheduled automation is an operations choice with data-scope implications, not a toy.

Voice Dictation and Read Aloud are documented in the same Learn unit; Microsoft states that real-time two-way voice chat requires a Microsoft 365 Copilot license. Name those controls only as Microsoft names them.

When techniques are not enough

Microsoft's Foundry guidance on customizing large language models treats prompt engineering, RAG, and fine-tuning as complementary, not a ladder you must climb for every idea. Microsoft Foundry's RAG article states the choice in leader language:

• Use RAG when you need answers grounded in private or frequently changing data.
• Use fine-tuning when you need to change model behavior, style, or task performance, rather than add fresh knowledge.
• Use agent tools (or another application pattern) when the system must call retrieval as a tool—or build a full RAG application when you need that control.

RAG retrieves passages (often through Azure AI Search in Foundry) and augments the prompt with that grounding data before generation. That is how you answer "is laser vision correction covered in this plan year?" against hundreds of pages of benefits PDFs without stuffing the entire corpus into every prompt by hand. RAG still uses prompts (system message: stick to retrieved content, cite sources). Poor chunking or retrieval still yields confident wrong answers. RAG also adds retrieval latency and extra input tokens. Microsoft is explicit that RAG quality depends on content preparation and that models can still fabricate despite grounding—enable citations and instruct the model to stay inside retrieved content.

Fine-tuning (supervised, in Microsoft's description) trains on example pairs to change behavior: a house style, a rigid output schema, or a task where the steering text is too long for the prompt window. Microsoft's illustrative case: natural language to SQL that was expensive and poorly grounded in schema became better and cheaper on a smaller fine-tuned model. Fine-tuning has training cost and often hosting cost even at idle. It is a poor substitute for a living knowledge base. Bad training data makes the base model worse—so you need a baseline before you celebrate.

A different tool. If the job is a refreshable dashboard, a deterministic payroll calculation, a workflow with approvals, or a system-of-record update, a chat prompt is the wrong product. Point the team to the application they already own (Excel, a line-of-business system, Power Automate, Microsoft Copilot Studio for an agent with governed knowledge) instead of stretching prompt poetry.

What a transformation leader does this week

• Publish a standard prompt card: goal, context, source, expectations, named reviewer.
• Stand up a gallery (Prompt Lab org collection or an internal library that feeds it).
• Train managers to iterate and to open sources.
• Draw a bright line: gallery prompts for drafting; RAG projects for living knowledge; fine-tune only after a baseline says behavior, not facts, is the gap; another tool when generation is not the job.
• Keep human review for anything that leaves the building.

The grounding and RAG chapter that follows this one goes deeper on business requirements for grounding. This chapter's job is to keep leaders from believing that a clever sentence is a knowledge-management strategy.

Official pages opened for this section

• Study guide for Exam AB-731 — skill Understand techniques of prompt engineering
• Prompt engineering techniques (Microsoft Foundry) — instructions, examples, cues, decomposition, grounding, recency, output structure
• Write an effective prompt — Copilot ingredients; /, #, and @ references
• Discover, share, and save prompts with Prompt Lab and Personalize Copilot with memory and custom instructions
• Customizing a large language model and RAG and indexes in Microsoft Foundry — when to use RAG vs fine-tuning vs prompt engineering