A prompt is the instruction and context you give an AI system. Good prompting is not about secret phrases. It is the practical work of making your goal legible: what you want, why you want it, what information should be used, what boundaries apply, and how the result should be presented.
The same principles apply to writing, analysis, images, and video, although each medium needs different descriptive details. This guide offers a repeatable method you can use in Cafa AI or another generative system.
The five useful parts of a prompt
Most strong prompts contain some combination of a task, context, audience, constraints, and output format. They do not need to be long. They need to remove the ambiguities that would most affect the result.
- Task: the action—explain, compare, draft, extract, generate, or revise.
- Context: background or source material the response should rely on.
- Audience: who will read, watch, or use the output.
- Constraints: length, tone, exclusions, required facts, or visual boundaries.
- Format: bullets, table, email, outline, scene description, or another explicit shape.
Weak versus improved prompts
‘Write about our launch’ leaves the purpose, audience, evidence, and format unresolved. A stronger version might read: ‘Draft a 150-word customer email announcing the feature described below. Address existing users, lead with the practical benefit, avoid hype, and end with one link placeholder. Do not add capabilities that are absent from the notes.’
Provide context without dumping everything
More context is helpful only when it is relevant. Label source material, distinguish instructions from quoted text, and explain which source takes priority if documents disagree. When an attachment contains untrusted instructions, ask the AI to treat it as data rather than as commands.
- Name the purpose of each file or excerpt.
- Include the section needed for the task rather than unrelated records.
- Remove personal, confidential, or regulated information unless its use is authorized and necessary.
- Ask the model to identify missing information instead of inventing it.
Use roles as perspective, not authority
A role can set vocabulary and point of view: ‘Explain this as a patient teacher’ or ‘Review this as an accessibility-minded editor.’ It does not grant expertise or make an answer correct. Avoid asking an AI to impersonate a real person or to replace a qualified professional in a high-stakes decision.
Specify a usable output format
Formatting instructions reduce cleanup. State the number of sections, desired headings, table columns, reading level, or aspect ratio where relevant. If a strict machine-readable format matters, request it explicitly and validate it after generation; generated syntax can still contain errors.
Iterate in small, observable steps
The first response helps reveal what the prompt omitted. Ask for a critique, select a direction, then revise. Preserve good elements explicitly. For image or video prompts, change a small number of attributes between generations so you can compare results instead of resetting the entire creative direction.
- First pass: establish structure or composition.
- Second pass: correct omissions and factual boundaries.
- Third pass: tune tone, pacing, style, or visual detail.
- Final pass: proofread and verify outside the AI system.
Common prompting mistakes
- Combining several conflicting goals in one request.
- Assuming the model knows private organizational context.
- Requesting ‘accurate sources’ without providing or checking sources.
- Using vague quality words such as professional or cinematic without concrete traits.
- Changing every variable at once during refinement.
- Pasting sensitive information that the task does not require.
A reusable prompt template
Use this as a starting point, not a ritual: ‘Help me [task]. The purpose is [goal], and the audience is [audience]. Use [sources/context]. Include [requirements] and avoid [exclusions]. Return the result as [format]. If essential information is missing, ask questions or label the uncertainty rather than filling the gap.’