The 10 Golden Rules of Prompt Writing
The 10 golden rules of prompt writing: goal, context, format and boundaries. Practical techniques that turn weak answers into strong results, with examples.
Ten rules are enough for good prompt writing: sharpen the goal, give the audience and context, separate the input data, write the boundaries and output format, show examples when needed, set a source rule, split complex work, test the result, and fix the prompt according to the cause of the error. These are not a magic formula. They are working rules that make the parts of the task invisible to the model visible.
If a prompt gave a good answer on the first try, that still doesn't show it works reliably. A different input, incomplete information or a changed model version can expose the weak spot. That's why in this article I tie every rule not only to "what do I write?" but also to "how do I check?"
The 10 golden rules of prompt writing
1. Write the outcome in one sentence first
"Help me with marketing" states a topic, not a job. "Extract the three main customer objections for the product page below and the proof each objection needs" already shows the outcome. If, when writing the sentence, you have no answer to "what will I do after receiving this response?", the goal is still foggy.
2. Show the audience and the place of use
The same topic does not speak the same language in an executive memo, a customer email and an Instagram post. Write who the reader is, what they already know about the topic and what step they should take after the text. A usage scene gives the model more information than "professional tone."
3. Select the necessary context
The model does not inherently know your company's earlier decisions, local terminology or budget limits. Provide the information that affects the decision; don't pour the whole archive into the prompt. Good context is not lots of information — it's the right information.
4. Separate the instructions from the input material
When adding long texts, email chains or assorted documents, create boundaries. Headings, triple quotes, Markdown blocks or a <document> tag can be used. OpenAI's prompting guide likewise recommends writing the instructions first and marking the context with clear separators.
5. Write the real constraints
Conditions like length, language, dates, legal limits, claims not to use and terms to preserve change the result. But stacking 30 prohibitions makes the answer brittle. Choose the three to five most important boundaries and, where possible, phrase them as "do this" rather than "don't do that."
6. Tie the output format to the decision
A table is not needed just because it looks nice. If you'll compare, show the columns; if a program will read it, show the fields; if an executive will read it, put a short verdict first. In an application that requires strict JSON, use the platform's schema-based structure feature instead of merely saying "return JSON."
7. Give examples as boundaries, not as decoration
If tone, classification or format is easier to show than to explain in words, add a few input–output examples. The examples should not cover only the comfortable cases. Anthropic's current guide says relevant, diverse and structured examples give steadier direction.
8. Set a rule for facts and unknown information
"Don't make mistakes" is not fact-checking. Define which claims will be verified against current sources, which sources take precedence, and what the model should write when information can't be found. For example: "If you can't find an official source, don't build a guess; list the missing information separately."
9. Split complex work into stages
When research, outline, draft, fact-check and final edit are demanded in the same answer, it becomes hard to see where an error occurred. Get the source plan first, then the structure, then the text, and finally the audit. But don't split a simple email into five stages; process is valuable only when it increases control.
10. Test the prompt with real inputs and version it
Choose at least three samples: typical, incomplete and risky. Record the results by accuracy, format, editing time and critical errors. Anthropic's prompt engineering overview requires defining a success criterion and a way to test it empirically before optimising the work.
All 10 rules applied in one prompt
"Prepare a project status from the meeting notes below. The text is written for the head of operations and the goal is choosing tomorrow's priorities.
Notes:
[add the text here]. First extract the confirmed decisions, then the delayed tasks, the responsible person, the date and the open questions. Don't mix proposals with adopted decisions.Give the answer in a five-column table: issue, status, owner, date, next step. If a note has no owner or date, don't invent one; mark it
to be clarified. At the end show only the three questions the manager must decide."
Here the goal, the reader, the input, the practical steps, the boundaries, the format and the unknown-information rule are in one place. Examples aren't needed because the output fields are shown specifically enough. If "proposal" and "decision" still get mixed up in testing, the next version can add one input–output example for each.
The practical difference between a weak and a strong prompt
| Weak version | Problem | More workable version |
|---|---|---|
| "Improve this text." | No criterion for "good" | "Without changing the meaning, delete repetition and internal jargon; edit it as a customer email under 150 words." |
| "Analyse the competitors." | Competitors, market and criteria unknown | "Compare [A] and [B] for [audience] by features, switching cost, support and data terms." |
| "Summarise the PDF." | The model picks what matters | "Extract the decision, three pieces of evidence, two limitations and the open questions with page numbers; don't write conclusions not in the document." |
| "Make the post more human." | No voice profile or fabrication boundary | "Follow the sentence rhythm of these three real samples; edit the draft without adding new facts or personal stories." |
| "Return correct JSON." | No field or type definitions | "Return an array with fields name:string, risk:low|medium|high, source:string|null; don't create extra fields." |
If you want 50 copyable variants of these examples, see the list of ready-made ChatGPT prompts. The point is not to keep the example as-is, but to see your own work's variables.
How do you fix a weak answer in three steps?
Name the error
Instead of "I didn't like it," write whether the problem is facts, tone, format, completeness or language. Pick the main error for the first fix.
Find the cause in the prompt
Was there no audience, no source, or did two constraints collide? If the problem is that the tool can't see current information, polishing the sentence won't be enough.
Update one variable and retest with the same sample
If you added examples, don't change the format and the goal at the same time. Keep the cause of the improvement visible.
A follow-up request example
"The structure of your answer is correct, but three claims are not in the given source. Remove those claims without adding new information.
Beside each remaining claim, write the page number supporting it. Move questions the source doesn't answer into a separate 'open information' section."
A 10-point check before sending a prompt
- Is the result I want clear in one sentence?
- Are the reader and the purpose of use visible?
- Is the context needed for the decision present?
- Are the instructions separated from the input material?
- Do the constraints avoid colliding with each other?
- Does the output format fit the next step of the work?
- Do the examples genuinely show the desired boundary?
- Is there a source and date rule for changing facts?
- Is the model's behaviour written for missing information?
- By which criteria will I judge the answer acceptable?
Not all of this list is needed for every simple request. Don't turn a two-sentence translation into a project document. Increase the control as the risk and workload grow. A broader structure and safety explanation is in the effective prompt guide.
Questions about prompt writing
Is assigning a role in the prompt important?
No. A role can narrow the viewpoint, but it doesn't replace the goal, the context and the evidence. Instead of "be an expert," write the specific responsibility and boundary.
Is it better to write prompts in English?
Not always. You can work in Azerbaijani. If technical terms get mixed up, give them with their English equivalent on first use and define the output language separately.
Does a long prompt give a more accurate answer?
Only if the extra words bring the context and criteria needed for the decision. Repetition, inflated roles and colliding prohibitions increase the length and can reduce the clarity.
How should the first answer be checked?
Check format and completeness first, then names, numbers, dates, quotes and sources. On high-risk topics keep a human expert and a primary source.
What separates a business prompt from a personal one?
In business work, acceptance criteria, confidentiality, ownership and reusability matter more. For sales, analysis and process examples, see the business prompts guide.
Sources
I'm Anar Rustamli - a strategist, entrepreneur, and AI adoption leader working at the edge of growth, technology, and human thinking. Since 2016, my work has focused on helping businesses evolve in a rapidly changing digital landscape. I design growth systems, AI-powered workflows, and strategic frameworks that align performance with purpose. I believe real growth happens when strategy, data, and human insight work together - and my mission is to help businesses adopt AI in a way that strengthens both their results and their identity.
