Microsoft AI-901: Prompting and Grounding Basics

Prompting is one of the foundational skills in AI-901. System instructions help define the model’s behavior, while user prompts specify the immediate task. Neither can supply reliable current or private business information by itself. When an application needs that information, grounding should provide evidence from approved sources and constrain answers to material the user is permitted to access.

The most useful distinction is simple: prompting tells the model what to do, while grounding gives it information it should use. Those two mechanisms work together but address different failure modes. If the instruction is vague, better documents may not help. If the instruction is clear but the facts are missing, better wording cannot invent reliable evidence.

System prompts define durable behavior

A system prompt establishes rules that should apply across many user requests. It can define the assistant’s role, tone, scope, safety boundaries, output format, and how to handle uncertainty. Keeping durable rules at this level prevents users from needing to restate them in every message.

System instructions should be specific enough to guide behavior but not overloaded with conflicting rules. Long prompts can become difficult to maintain, especially when multiple teams add requirements. Good prompt engineering includes removing redundant instructions and testing whether the model actually follows the priorities that matter.

User prompts should state the immediate goal clearly

A good user prompt explains the task, provides relevant inputs, and states the expected output. “Summarize this” is less useful than “Summarize this incident report in five bullets, preserving dates, affected systems, and unresolved actions.” The second prompt defines what a successful answer looks like.

Examples can help when a desired format is difficult to describe. However, examples also consume context and can bias the model toward patterns that do not fit every case. Use them when they clarify a real requirement rather than adding them automatically.

Prompt structure reduces ambiguity

Separating instructions, source material, and output requirements helps the model interpret the request. Delimiters or clear labels can distinguish quoted evidence from commands. This becomes especially important when the source content itself may contain instructions, such as documents or web pages processed by an AI application.

Structured prompts also improve maintenance. Teams can update one instruction block without accidentally changing the data section. The goal is not elaborate formatting for its own sake; it is reducing ambiguity between what the model should do and what it should merely read.

Grounding supplies facts the model may not know

Grounding is useful for company policies, product catalogs, current procedures, support knowledge, or other information that is private or changes over time. The application retrieves relevant material and provides it to the model as context for the response.

This pattern is powerful because it avoids expecting the model’s training to contain every business fact. The model can focus on interpreting and explaining evidence. The broader Azure AI application path increasingly relies on this separation between a general model and application-specific context.

Retrieval quality determines grounding quality

If the retrieval step selects the wrong documents, the model may answer confidently from the wrong evidence. Applications need sensible chunking, indexing, metadata, search logic, and ranking. The correct answer might exist in the knowledge source but still never reach the model.

Evaluation should therefore inspect both stages. First ask whether the retrieval system found the right evidence. Then ask whether the model used that evidence correctly. Combining the two into one score can hide the root cause of poor answers.

Grounded answers should distinguish evidence from inference

In high-trust use cases, users should be able to tell which parts of an answer are supported by sources and which parts are the model’s interpretation. Citations or source references can make verification easier, particularly when policies, technical procedures, or regulated information are involved.

If the retrieved material is incomplete, the model should not be encouraged to fill the gap with plausible language. A good instruction can require the model to say when evidence is insufficient. That behavior is often more valuable than producing an answer to every prompt.

Prompt injection is a boundary problem

When an application retrieves documents or processes user-supplied content, that content may contain text that looks like instructions. The application should treat external content as data, not automatically as trusted control text. Clear separation between system rules and retrieved material reduces the chance that untrusted content changes the model’s behavior.

Permissions matter too. Grounding should not retrieve documents the user is not allowed to access. Natural-language search can make sensitive information easier to discover, so the retrieval layer must enforce authorization before content reaches the model.

Evaluate prompts with a stable test set

Prompt improvement should be measured rather than judged from one impressive response. Keep a set of representative requests and expected behaviors, then compare prompt versions against the same cases. Include normal tasks, edge cases, ambiguous input, and unsupported questions.

Track whether the model follows instructions, uses the evidence, returns the required format, and avoids unsupported claims. A prompt change that improves style but reduces factual grounding is not necessarily an improvement.

Practice the complete Foundry interaction, not isolated wording tricks

The current exam expects candidates to deploy a model and interact with it in Microsoft Foundry, not merely define prompt vocabulary. Build a small exercise where a system prompt sets rules, a user prompt asks a task, and supplied context grounds the answer. Change one part at a time and observe what happens.

Within Microsoft certifications, AI-901 establishes how prompts, sources, and evidence shape model behavior; Microsoft AI-300 later examines the operational systems used to deploy and monitor advanced workloads. AI-901 remains the foundation: understand how instructions and evidence shape model behavior before moving into more complex AI engineering.

Prompting should not become a ritual of copying popular phrases. The best prompt is the one that makes the application’s requirements clear and produces reliable behavior on its real workload. Grounding follows the same rule: retrieve only evidence that materially helps answer the question. Clear instructions plus relevant evidence is a stronger design than either one used alone.

Grounding sources should also be curated. A retrieval system that indexes obsolete policies, duplicate documents, or conflicting drafts can create more confusion than value. Source quality, versioning, and metadata are part of the AI solution even though they are not model features. Good grounding begins with trustworthy content.

Prompt templates can help teams maintain consistency, but they should not become rigid copies used for unrelated tasks. A support summarization prompt, a data-extraction prompt, and an agent instruction have different goals. Reusing the same long “universal prompt” can hide requirements and make testing harder.

Few-shot examples are useful when the model needs to imitate a particular format or decision pattern. One or more examples can show the desired behavior more clearly than a long explanation. However, examples should be representative and should not contain accidental biases that cause the model to overfit the pattern.

Applications should also protect against context poisoning. If retrieved material can be edited by untrusted users, an attacker may insert misleading text that the model later treats as evidence. Authorization, source trust, content review, and clear instruction hierarchy reduce that risk. Grounding is only as trustworthy as the data pipeline behind it.

Another good practice is to separate factual evaluation from style evaluation. A response can be well written and still unsupported. Score whether the answer uses the correct evidence before judging tone or fluency. This helps teams avoid optimizing prompts for polish while factual quality quietly degrades.

When grounding is unavailable, the prompt can define fallback behavior. The model might ask the user for more information, say that it cannot verify the answer, or provide a general explanation while clearly labeling it as such. Designing that behavior in advance is safer than letting the model improvise when evidence is missing.

Grounding can also support freshness. When information changes frequently, updating the source repository is usually more practical than expecting a model’s training to contain the latest version. This makes retrieval especially useful for internal policies, product documentation, support procedures, and other living knowledge bases.

For hands-on practice, take one question whose answer exists in a short document. Ask the model without the document, then provide the document as grounding context and compare the result. Next, make the user prompt ambiguous and observe whether clearer instructions improve the answer. That simple experiment demonstrates the separate value of evidence and prompting.

The broader generative AI learning path can provide more depth, but AI-901 preparation should stay anchored in the current Foundry skills: effective prompts, deployed models, lightweight applications, and agents. Grounding is most useful when it supports those implementation patterns rather than becoming an abstract retrieval topic.

Keep prompts maintainable as the application evolves. Store important prompt versions, test changes against the same evaluation cases, and avoid editing production instructions without evidence. Even a small wording change can affect formatting, refusal behavior, or how the model uses grounded context.

Versioned prompts also make incidents easier to reproduce because teams can identify exactly which instructions produced a problematic response. Pair prompt-version evidence with model settings and retrieval inputs when investigating regressions, because a prompt alone rarely explains the whole response.