Think of a prompt as a brief for a capable assistant
Modern AI models can understand natural language, files, images and long instructions, but they still need to know what outcome you want. Vague prompts often produce generic answers because the model has to guess your goal, audience, constraints and preferred format.
Five parts of a strong prompt
Task
State the action clearly: compare, summarize, draft, analyze, rewrite, troubleshoot, plan or critique.
Context
Explain the audience, situation, purpose and background the model needs to understand the task.
Inputs
Provide the source text, data, examples, links, files or facts the answer should use.
Constraints
Specify length, tone, exclusions, must-use details, date range, geography, assumptions or other boundaries.
Output format
Ask for the structure you need: table, email, checklist, code, JSON, short paragraphs, comparison or step-by-step plan.
Write something about email marketing.
Write a 500-word beginner guide to email marketing for small local businesses. Explain list building, subject lines and one simple welcome sequence. Use short paragraphs, avoid jargon, and finish with a five-point checklist.
Use TCREI to write and improve a prompt
The original version of this article introduced the TCREI framework. It remains useful because it combines prompt construction with an explicit review loop.
| Step | Meaning | What to do |
|---|---|---|
| T | Task | Define exactly what the AI should produce or do. |
| C | Context | Provide audience, purpose, background and constraints. |
| R | References | Give examples, source material or a structure to follow. |
| E | Evaluate | Check accuracy, relevance, tone, completeness and formatting. |
| I | Iterate | Refine the instructions or ask for a targeted revision. |
Let the AI interview you before answering
You do not always know the perfect prompt in advance. For planning, strategy, writing or technical tasks, ask the model to collect missing information first.
Interview me first. Ask one question at a time. When you have enough information about my goal, audience, constraints and available resources, summarize your understanding and then give me the final recommendation.
This is especially useful for a presentation, content brief, project plan, travel plan, software requirement or any task where an important missing detail could change the answer.
Show the model what “good” looks like
If format or style matters, include one or more examples. This is often called few-shot prompting. A short example can communicate tone, structure and level of detail more precisely than several paragraphs of abstract instructions.
Required style: concise product descriptions with one benefit-led sentence followed by three bullets.
Reference: “Compact desk lamp for focused work. • Adjustable brightness • USB-C power • Foldable design.” Now write the next five products using the same structure.
Separate context, instructions and output requirements
For long prompts, use headings or clear delimiters so the model can distinguish your source material from your instructions.
# Context
[Paste the background or source material]
# Task
Analyze the material and identify the three most important issues.
# Constraints
- Do not invent missing facts.
- Separate facts from assumptions.
- Keep the answer under 500 words.
# Output
Return a table with Issue, Evidence and Recommended Action.When you provide a large document or dataset, place the source material first and the exact question after it. This makes the final instruction easier to identify.
Use a role when it adds useful perspective—not as a magic formula
“Act as a…” can help define the expected perspective, but a role alone is not enough. “Act as an SEO expert” is much weaker than specifying the page, target audience, data sources, business goal, constraints and expected output.
You are reviewing this landing page as a technical SEO specialist. Use the supplied Search Console export and page HTML only. Identify indexation, metadata and internal-linking issues. Separate confirmed issues from recommendations, and rank the fixes by effort and likely impact.
Ask AI to challenge your assumptions
A leading prompt can push an AI system toward the answer you appear to prefer. For decisions, analysis and advice, make the prompt neutral and explicitly ask for counterarguments.
Leading
“This is clearly the best strategy for our website. Don’t you agree?”
Better
“Evaluate this strategy objectively. Identify strengths, weaknesses, missing evidence and at least two alternative approaches.”
For a detailed explanation of this behaviour, read AI Sycophancy: Why AI Agrees With You.
Prompt for evidence when facts matter
Better wording cannot guarantee a correct answer. When the task depends on current facts, calculations, legal rules, product specifications or other verifiable information, ask the model to use appropriate tools or sources when available—and check important claims independently.
Do not assume my interpretation is correct. Distinguish verified facts from assumptions. For time-sensitive claims, use current sources if web access is available and cite them. If the available evidence is insufficient, say what information is missing.
Give AI the data and ask it to show its work
When working with code or data, provide the actual error, sample rows, schema or file whenever possible. Ask for the method and checks—not only the final answer.
Using only the uploaded CSV, calculate median salary by experience band. Show the cleaning steps, row counts before and after filtering, the code used for the calculation, and one chart. State any assumptions or limitations.
See the practical workflow in Google Colab Data Science Agent.
Four reusable prompt templates
Writing
Draft [content type] for [audience] with the goal of [goal]. Use [tone]. Include [must-have points]. Avoid [exclusions]. Keep it to [length]. Return it as [format].
Research / comparison
Compare [A] and [B] for [use case]. Use these criteria: [criteria]. Separate documented facts from interpretation, identify missing information, and present the result in a table followed by a short explanation.
Critique
Review the material below. Do not assume it is correct. Identify factual gaps, weak reasoning, unclear sections and unsupported claims. Then propose specific revisions without changing parts that already work.
Problem solving
Help me solve [problem]. First summarize the symptoms and constraints. Generate the three most plausible causes, explain how to test each one, and recommend the next action based on the evidence I provide.
Prompting mistakes that reduce answer quality
- Asking several unrelated tasks in one large prompt.
- Providing no audience, goal or output format.
- Using persuasive or leading wording when you want an objective assessment.
- Giving examples that conflict with the written instructions.
- Asking for “the latest” without giving the model access to current information.
- Treating a fluent answer as automatically accurate.
- Overloading a simple task with unnecessary role-play or complex frameworks.
- Failing to evaluate and refine the first result.
Watch the AI prompting walkthrough
The video demonstrates the same core ideas: provide context, state the task clearly, specify the output and improve the prompt through interaction.
