A prompt is the instruction a practitioner gives to an AI tool. The quality of the output depends substantially on the quality of the prompt, and this reflects the central skill of working with AI rather than an incidental technical detail. A practitioner who sends an AI tool a vague request gets vague output. A practitioner who sends a clear, structured instruction gets output that often comes close to usable on the first attempt and that can be refined quickly.
The most useful mental model for a prompt is that it is a brief. The same way a senior would brief a junior consultant before assigning them a piece of work, the practitioner briefs the AI before expecting useful output. A good brief contains five components: who is doing the work, what needs to be done, what the situation is, what the output should look like, and what rules apply. A prompt that contains these components produces substantially better results than a prompt that does not, and the discipline of including them becomes habitual with practice.
Role
The role tells the AI who it should be in producing the response. An AI tool without a specified role tends to produce outputs in a neutral, averaged voice drawn from the full range of writing in its training data. A specified role pulls the output toward a specific register, specific vocabulary, and specific depth of treatment. "You are a senior associate at a corporate law firm preparing a memo for a partner" produces different output from "You are a legal educator explaining this to a student." Both roles produce competent legal content. They produce competent legal content of different kinds, suited to different purposes, and the practitioner gets whichever kind they specified.
A good role instruction combines professional identity, seniority level, and audience. Professional identity tells the AI what domain to operate in. Seniority level tells it how much to assume and how much to explain. Audience tells it how to pitch the output. "You are a senior risk manager at a European bank briefing non-technical board members on cyber risk" gives the AI enough information to produce a board-appropriate treatment. "Explain cyber risk" does not, and the output reflects the difference.
The role is also the fastest single adjustment a practitioner can make when an output is not right. If the first response is too casual, too technical, too generic, or too detailed, changing the role and resubmitting often produces a substantially better result with no other change. This is worth internalising. A single role adjustment frequently outperforms more elaborate rewrites of the rest of the prompt.
Task
The task tells the AI what action to take. A good task instruction uses specific verbs that identify the cognitive operation the practitioner wants performed. "Summarise" is a specific verb. "Tell me about" is not. "Critique" is specific. "Explain" can be specific with qualification. "Draft," "compare," "evaluate," "extract," "restructure," "propose alternatives," "analyse the assumptions in": these are specific task verbs that tell the AI what kind of work to do.
A weak task instruction describes a topic. A strong task instruction describes an operation on a topic. "The company's Q3 results" is a topic. "Summarise the three most important points from the company's Q3 results for a busy executive" is an operation. The first produces generic output. The second produces usable output.
Task instructions can also be layered. A single prompt can contain multiple sequenced tasks, which often produces better results than a single compound task. "First, summarise the main points from this document. Second, identify the three assumptions that drive the author's conclusions. Third, propose two questions that would test whether those assumptions hold." This pattern works because it gives the AI a clear sequence rather than a single request that bundles multiple cognitive operations together. The layered pattern scales well; a practitioner can include four or five sequenced tasks in a single prompt as long as each is specific.
Context
The context tells the AI what situation the task exists within. Without context, the AI produces generic output that could apply to any situation in its training data. With context, the AI can tailor the output to the specific situation the practitioner is working in. Context includes the source material the AI should work with, the organisational reality that shapes what is possible, the objectives that define what success looks like, and the constraints that bound the problem.
Source material is often the most valuable context. A practitioner asking an AI to draft a memo on a complex matter produces better output when the AI has the underlying documents than when the AI is working from the practitioner's summary alone. Pasting the relevant source material into the prompt or attaching documents through whatever mechanism the tool supports provides the AI with the specific information it needs to produce specific output. The tool's context window (covered in Module 1.2) determines how much source material can be provided in a single interaction, and the practitioner benefits from providing the material that is most relevant to the specific task rather than everything that exists.
Organisational context is often underused. A consultant asking an AI to propose a restructuring approach for a client gets more useful output when the prompt mentions that the client is a mid-market European services firm with 400 employees and strong cost pressure than when the prompt asks for generic advice. Context does not need to be long. A sentence or two identifying the key features of the situation often produces substantial improvement in the relevance of the output.
Format
The format tells the AI what the output should look like. Without format guidance, the AI produces output in whatever shape its training data suggests is most common for the kind of request. With format guidance, the output arrives in a shape the practitioner can use directly.
Format instructions can specify the type of document (memo, email, executive summary, bullet list, comparison table, decision matrix), the length (one paragraph, 200 words, two pages), the structure (heading hierarchy, section order, use of bullet lists or prose), and the level of detail (high-level overview, detailed treatment with specific examples). "Produce a one-page executive summary with three main findings and two recommendations, each with a single paragraph of supporting rationale" tells the AI what the practitioner will accept. A prompt without format guidance produces output the practitioner may then have to restructure, which adds time rather than saving it.
Constraints
Constraints tell the AI what rules apply. Constraints include what to avoid, what standards to maintain, what to omit, and what to treat as fixed. "Do not include content that is speculative beyond what the source material supports" is a constraint. "Write in active voice" is a constraint. "Use the firm's style guide for numerical formatting, which treats thousands with commas and percentages to one decimal place" is a constraint. "If specific information is not provided, mark the location and describe what information would be needed" is a constraint.
Constraints work best when they are specific to the situation rather than generic. A practitioner who says "be professional" provides less useful constraint than one who says "maintain a formal register suitable for client communication in a regulated industry, avoid casual phrasing, do not use contractions." The AI can act on specific constraints. Generic constraints get averaged into the default behaviour.
Putting the Components Together
A prompt that combines all five components produces substantially better output than a prompt that uses only one or two. A worked example illustrates the difference. Consider a consultant who needs to produce a briefing note for a client board on a proposed market entry.
A weak prompt: "Write me a briefing note on entering the Polish market."
A strong prompt containing all five components: "You are a senior consultant at a global management consulting firm [role]. Produce a briefing note on market entry into Poland for the board of a client company [task]. The client is a mid-market UK industrial services firm with 300 employees and established operations in the UK and Germany, exploring Central European expansion; the board meets in three weeks and has asked for an initial view before commissioning fuller work [context]. Format the output as a two-page briefing note with the following sections: executive summary, market overview, entry options, key risks, recommended next steps [format]. The treatment should reflect current market conditions, identify where assumptions would need to be tested with further work, avoid specific financial claims that would require verified sources, and flag any regulatory considerations that would materially affect the decision [constraints]."
The strong prompt is longer to write. The output it produces is typically usable with modest editing, while the weak prompt produces generic output that would need substantial revision before it could be presented to anyone. The time invested in writing the prompt is recovered several times over in reduced editing time, and the quality of the final output is typically better because the structure of the request prompts the AI to produce structured thinking.
The investment in writing good prompts compounds over time. A practitioner who develops the habit of including all five components produces better output faster, and the habit becomes automatic after a few weeks of deliberate practice. Practitioners who have been using AI tools for a year or more often report that they barely think about the five components explicitly because the pattern has become part of how they formulate any AI request.
Why Most Prompts Fail
Three failure patterns account for most weak prompts. Vagueness is the most common. The practitioner sends a request that does not specify the role, task, context, format, or constraints clearly enough for the AI to produce focused output. Overloading is the second. The practitioner packs too many cognitive operations into a single prompt, producing output that handles each operation shallowly rather than any of them well. Under-contextualisation is the third. The practitioner does not provide enough of the specific situation for the AI to produce situation-specific output, and the result is generic material that could apply anywhere.
Recognising these failure patterns in one's own prompts is a useful diagnostic when output disappoints. A prompt that produces weak output has typically failed on one of these three dimensions. Adjusting the prompt to address the failure pattern usually produces substantially better output on the next attempt, which is why iteration (covered in Section 5) is central to productive AI use.