Most disappointing AI results come from the same cause: the instruction was too thin. “Write a blog post about email marketing” gives an assistant almost nothing to work with. It does not know who the post is for, how long it should be, what you already believe, or what a good result looks like. So it guesses, and the guess is usually generic.
Prompting is not a trick or a secret language. It is closer to briefing a capable colleague who has never met you and cannot ask follow-up questions unless you invite them. This guide gives you a framework for writing that brief, shows how each part changes the result, and ends with templates and a troubleshooting list you can use right away. The principles apply to any text-based AI assistant, so they will not go out of date when products change.
Key takeaways
- A good prompt states the goal, the context, the audience, the format and the constraints. Most weak prompts leave out two or three of these.
- Examples teach faster than descriptions. Showing one sample of what you want often beats a paragraph of explanation.
- Treat the first answer as a draft. Iterate with specific feedback instead of starting over.
- Ask the assistant to flag uncertainty and list assumptions, then check facts yourself.
- Save prompts that work. A small library of tested templates saves more time than any single clever phrase.
What a prompt really is
A prompt is the full set of information you give the assistant before it responds. That includes your question, but also any background, files, examples and instructions you add. The assistant generates its answer based on all of it, and only that. It does not know your business, your preferences or your history unless you tell it, or unless the tool has a memory or settings feature that does.
This has a practical consequence. If you hold information in your head and leave it out of the prompt, the assistant cannot use it. Better prompts are usually not cleverer. They are more complete.
Why vague prompts produce generic answers
When an instruction is open-ended, the assistant fills the gaps with the most common choices. Ask for “a marketing email” and you get the average marketing email: friendly opening, three benefits, a call to action. It is not wrong. It is just nobody’s email in particular. Every detail you add narrows the range of possible answers toward the one you actually want.
The GCAFC framework
A simple way to remember what to include is five parts: Goal, Context, Audience, Format, Constraints. You do not need all five every time, but when a result disappoints, one of them is usually missing.
1. Goal: what you want to achieve
State the outcome, not only the task. “Write a product description” is a task. “Write a product description that helps a first-time buyer decide between our two models” is a goal. The second tells the assistant what the text has to accomplish, which shapes every choice it makes.
2. Context: what the assistant needs to know
Include the background that a smart outsider would need. Who you are, what the product or project is, what has already happened, and any facts the answer must respect. If you have source material such as notes, a draft or data, paste it in and say how to use it. A useful habit is to separate instructions from material using clear labels, for example “Instructions:” and “Source text:”, so the assistant does not confuse the two.
3. Audience: who will read or use the result
The same information reads very differently for a chief financial officer and for a college student. Name the reader, their level of knowledge and what they care about. “Explain this for a busy small business owner with no technical background” produces a different answer than “explain this for a software engineer”.
4. Format: what the output should look like
Say what shape you want: a table, a numbered list, three short paragraphs, a headline with five options, an email with a subject line. Give a length where it matters, such as “about 150 words” or “no more than five bullet points”. If you plan to paste the result somewhere, say so, and request plain text or specific formatting accordingly.
5. Constraints: what to include, avoid or follow
Constraints are the guardrails. Tone of voice, words to avoid, claims you cannot make, sources you must use, topics that are out of scope, reading level. Constraints stop the assistant from drifting into a style or content you would have to remove later. Positive instructions usually work better than negative ones, so “write in short, plain sentences” tends to work better than “do not be wordy”.
Before and after: three worked examples
Example 1: A marketing email
Weak prompt:
Write an email about our new feature.
Stronger prompt:
Goal: Get existing customers to try our new scheduling feature this week.
Context: We run an online booking tool for yoga studios. The new feature lets clients reschedule a class themselves in two taps. Studio owners told us they spend a lot of time handling reschedule requests by message.
Audience: Studio owners who already use us, are busy and are not technical.
Format: A short email with a subject line, three short paragraphs and one clear call to action button text.
Constraints: Friendly and direct. Under 150 words. Do not claim specific time savings. Do not use exclamation marks.
The second prompt gives the assistant a reader, a problem to speak to and boundaries. The result will still need your review, but it starts much closer to usable.
Example 2: Summarising a document
Weak prompt:
Summarise this report.
Stronger prompt:
Summarise the report below for a director who has two minutes to read it. Start with the single most important finding. Then list the three decisions the report asks for, with the deadline for each. Finish with any risks the authors flag. If something is unclear or missing, say so instead of guessing. Report: [paste text]
Notice the last instruction about unclear or missing information. It gives the assistant permission to be honest instead of filling gaps, which reduces made-up details.
Example 3: Brainstorming
Weak prompt:
Give me ideas for a webinar.
Stronger prompt:
I run a bookkeeping service for freelancers. Suggest 10 webinar topics that would attract freelancers who have never used accounting software. For each topic, give a working title, the main problem it solves and one concrete thing attendees will leave with. Avoid topics about tax law changes. Present the results as a table.
The stronger prompt names the business, the audience, the output structure and an exclusion. You will still choose which ideas are good, but you will have ten relevant ones to choose from.
Use examples to show, not tell
Describing a style is hard. Showing it is easy. If you have a sample of the tone, structure or format you want, include it. This technique, often called few-shot prompting, is one of the most reliable ways to improve consistency.
How to do it well
- Give one to three examples. One is often enough. More can help when the pattern is subtle.
- Label them. Mark each as “Example input” and “Example output” so the pattern is clear.
- Choose representative samples. If every example is short and cheerful, expect short and cheerful results.
- Say how closely to follow them. “Match the structure and tone, but do not reuse any phrases” prevents copying.
A short example
Rewrite product notes into customer-friendly release notes. Follow the style of this example.
Example input: “Fixed null pointer on export when no filters set.”
Example output: “Exporting without filters now works as expected.”
Now rewrite these notes: [paste notes]
The example teaches the transformation more clearly than a paragraph of explanation could.
Roles and personas: when they help
Many people start prompts with “You are an expert in…” It can help, but less than people think. A role works best when it signals a perspective, a level of detail and a way of reasoning that the task needs, such as “Act as a careful editor who checks for logic gaps” or “Act as a skeptical buyer reading this landing page”.
A role does not give the assistant knowledge it lacks, and “world-class expert” language adds little. The useful part is the specific perspective. If a role is not changing what you want the assistant to do, leave it out and spend those words on context and constraints instead.
Break big tasks into steps
Large requests such as “write a complete report” tend to produce shallow results because everything is attempted at once. Splitting the work into stages gives you control and usually raises quality.
A staged approach to a long article
- Clarify. Ask the assistant what else it needs to know before it starts. Answer its questions.
- Outline. Request a structure with headings and a one-line summary for each section. Edit it.
- Draft section by section. Ask for one section at a time, referring to the approved outline.
- Review. Ask the assistant to critique the draft against your goal and audience, then revise.
- Polish. Final pass for tone, length and consistency, which you do or check yourself.
This approach also makes it easier to catch problems early. It is far cheaper to fix an outline than a finished document.
Ask for reasoning when accuracy matters
For problems that involve several steps, such as calculations, comparisons or planning, ask the assistant to work through the steps and show its reasoning before giving the final answer. Reviewing the steps lets you spot where it went wrong. You can also ask it to double-check its own result, but treat that as a helpful aid, not a guarantee.
Iterate with specific feedback
The first response is a draft. The skill is in the follow-up. Vague feedback like “make it better” gives the assistant nothing to act on. Specific feedback gets results.
| Instead of | Try |
|---|---|
| “Make it better.” | “Shorten the second paragraph by half and lead with the benefit.” |
| “Sounds too robotic.” | “Use contractions and shorter sentences. Replace formal words like ‘utilise’ with plain ones.” |
| “Not what I wanted.” | “The audience is beginners, so remove the jargon and add one everyday analogy.” |
| “More creative.” | “Give five alternative openings, each using a different approach: a question, a statistic placeholder, a short story, a bold claim, a contrast.” |
When a conversation has drifted, it can help to start a fresh one with a cleaner prompt that includes everything you have learned. Long threads sometimes carry along earlier mistakes and conflicting instructions.
Reduce mistakes and made-up facts
AI assistants can state false information confidently. Prompt design cannot remove this risk, but it can lower it.
- Provide the source material. Ask the assistant to answer only from the text you supply, and to say when the text does not contain the answer.
- Allow uncertainty. Add “If you are not sure, say so” and “list any assumptions you made”.
- Use placeholders for facts you have not supplied. For example, “Where a statistic would help, insert [STAT NEEDED] instead of inventing one”.
- Ask for sources, then verify them. Assistants can produce references that look real but are not. Check every citation yourself before using it.
- Review high-stakes output. Anything involving legal, medical, financial, safety or published claims needs a qualified human check.
Treat every important claim as unverified until you have checked it against a reliable source.
Ready-to-use prompt templates
Copy these, replace the bracketed parts and adjust to your needs.
General-purpose brief
Goal: [what this should achieve]
Context: [background and facts the answer must respect]
Audience: [who will read or use it, and what they already know]
Format: [structure, length, layout]
Constraints: [tone, things to include, things to avoid]
Before you start, ask me up to three questions if anything is unclear.
Rewrite for clarity
Rewrite the text below for [audience] at a [reading level] level. Keep the meaning and all facts unchanged. Make sentences shorter, replace jargon with plain words and keep the same length or shorter. List any places where you were unsure about the meaning. Text: [paste]
Critique my draft
Act as a careful editor. Review my draft against this goal: [goal]. Give feedback in three groups: what works, what is unclear or weak, and what is missing. Be specific and quote the lines you are referring to. Do not rewrite the draft yet. Draft: [paste]
Compare options
Compare [option A] and [option B] for [specific use case]. Use a table with these criteria: [list]. Then give a recommendation for someone who values [priority]. Mark anything you are not certain about so I can verify it.
Turn notes into a plan
Turn the notes below into an action plan. List each action with an owner placeholder, a suggested deadline and any dependency. Group actions by theme. Flag anything that looks ambiguous. Notes: [paste]
Troubleshooting weak results
| Problem | Likely cause | Fix |
|---|---|---|
| Generic, bland output | Missing context or audience | Add specifics about the reader, the situation and what makes this case different |
| Wrong tone | Tone not defined | Describe the tone and paste a sample of writing that sounds right |
| Too long or too short | No length given | State a word count or section count |
| Ignores part of your instructions | Instructions are buried or conflicting | Use a numbered list, put the key rule first and remove contradictions |
| Invented facts | No source material, no permission to say “unknown” | Supply sources, allow uncertainty and verify claims |
| Inconsistent across runs | Open-ended prompt | Add an example and a fixed format |
| Rambling answers | No structure requested | Specify format and ask for the main point first |
Build a prompt library
The biggest productivity gain from prompting comes from reuse. When a prompt works, save it with a short note on what it is for, what to replace and what a good result looks like.
- Keep it simple. A shared document or notes app is enough. Organise by task, such as “email replies”, “meeting summaries” and “research briefs”.
- Record versions. When you improve a prompt, keep the old one for a while so you can compare.
- Add an example output. A good sample helps colleagues know what to expect.
- Review quarterly. Remove prompts nobody uses and update the ones that no longer fit.
- Note the data rules. Mark which prompts must never include customer or confidential information.
Frequently asked questions
How long should a prompt be?
As long as it needs to be to include the goal, context, audience, format and constraints. Many effective prompts run to a short paragraph or a short list. Length is not the aim. Completeness is. If extra words do not change the answer, remove them.
Do magic phrases like “take a deep breath” or “this is very important” work?
Occasional tricks may nudge results in some cases, but they are unreliable and vary between tools. Clear instructions, relevant context and examples work consistently and are worth your attention first.
Should I say please and thank you?
Politeness does not hurt, and some people find it keeps their own writing clear and respectful. It is not a reliable way to improve results. Clarity matters more than courtesy words.
Why do I get different answers to the same prompt?
These systems generate text with a degree of randomness, so outputs vary. To get more consistent results, make the prompt more specific, include an example of the format and ask for a fixed structure.
Can I paste confidential information into a prompt?
Only if you have confirmed the tool’s data handling terms and your organisation allows it. Until then, remove names, numbers and sensitive details or replace them with placeholders.
How do I know if my prompt is good?
Test it on three different inputs. If the results are useful each time with minimal editing, it is good. If one input fails, look at what was different and add that case to your constraints or examples.
Conclusion
Better prompts come from better briefs. State the goal, give the context, name the audience, specify the format and set constraints. Show an example when style matters, split large jobs into stages and give specific feedback when you iterate. Allow the assistant to say it is unsure, and always check what matters before you rely on it.
Start with one task you do every week, write a complete prompt for it, test it three times and save the version that works. Within a month you will have a small library that does more for your productivity than any single trick. To choose the right assistant for your work in the first place, see our checklist for choosing an AI tool.
