How to Write Better Prompts for AI
23 Jun 2026 · RS Management
TL;DR
- The result depends on the context in the prompt, not on magic phrases.
- A prompt worth reusing has five parts: role, context, task, output format and an example of a good answer.
- On longer tasks, fixing the plan takes 2 minutes and fixing the finished document takes 30.
- Every defect in an answer is a missing instruction to add to the prompt.
“How is AI supposed to save me time if after three attempts I still fix everything by hand?” The difference between a person who gets useful results from a model and one who gives up after a few tries rarely comes from knowing special phrases. It comes from how much information goes into the prompt and how it is organized. Lists of magic phrases such as “act as a world-class expert” add no knowledge about the task. Across 2,410 factual questions and 162 roles written into the prompt, adding the role alone did not improve answer accuracy compared with no role.1
Applies to ChatGPT Claude Gemini Copilot
Five parts of a good prompt
| Part | What it contains | Example |
|---|---|---|
| Role | Who the model is and who it writes for | “a financial analyst preparing material for the board” |
| Context | What the model needs to know so it does not guess: industry, project stage, constraints, earlier agreements | “three years of cooperation, a framework agreement with an annual annex” |
| Task | One action expressed as a verb | “list the risks”, and in a separate step “assess which one is the largest” |
| Output format | Length, structure, tone, language, table layout | “up to 150 words, no exclamation marks” |
| Example | One or two patterns of a good answer | two earlier emails to the same recipient |
Context should include what is needed, not everything available. Models get distracted by irrelevant fragments and make more mistakes on tasks padded with extra, unneeded data.2
Three pairs from everyday work
Meeting notes
Before:
Summarize these meeting notes.
After:
You are a project manager’s assistant. Below are the raw notes from a 90-minute status meeting with the client and two suppliers. Prepare a summary for a director who was not at the meeting and has two minutes to read it. Format: three sentences with the key points, then a list of decisions taken, then a table of tasks laid out as task / owner / deadline. If an owner or deadline was not stated explicitly, write “not set” instead of filling it in with your own guess.
The last sentence removes the most common defect of such summaries: deadlines nobody agreed on.
Replying to a price increase
Before:
Write an email to the supplier saying we don’t accept the price increase.
After:
Context: three years of cooperation with a maintenance service supplier, a framework agreement with an annual annex, the supplier has announced a 12% rate increase from January and justifies it with labor costs. We want to keep working together and to split the increase into two stages. Write a reply of up to 150 words, in a matter-of-fact and calm tone, without exclamation marks, ending with one question about willingness to discuss the schedule. Below are two of our earlier emails to this supplier; keep to that tone.
Two pasted emails work better than any description of style. A model reproduces a pattern well and guesses intent poorly.
Analyzing a sales sheet
Before:
Analyze this sales data.
After:
Below is monthly sales for 14 product categories over the last 18 months. Task: identify the categories where year-on-year growth changed direction in the last quarter. For each, give the size of the change and the month the trend reversed. Format: a table with the columns category / year-on-year change / month of reversal / short comment. Do not interpret the causes. If data for a category is incomplete, list it separately below the table.
The sentence “do not interpret the causes” has a practical purpose. An analysis request without limits ends with explanations that cannot be derived from the numbers provided.
A plan before execution
For tasks longer than one page, ask for a plan first: 5 to 7 points saying what will go into each section and where the data will come from. The plan also exposes a misunderstanding at the start, when the model has read the task differently from what was intended. This habit pays off most on requirements documents, proposals and analyses.
Iterating on your own errors
After a weak answer, the first reflex is to fix it by hand. It pays more to treat each defect as a missing instruction and add it to the prompt:
| Defect in the answer | Instruction added to the prompt |
|---|---|
| Generalities instead of numbers | A requirement to give a numerical value with every claim |
| Three pages instead of one | A hard word limit |
| Own conclusions instead of facts | “Do not interpret, give only what follows from the data” |
| Missing data filled in with a guess | “If the information is not in the material, write ‘no data’” |
After a few rounds you get a prompt that produces a predictable result on the first try. Proven prompts are best kept in one file or in a team library, grouped by category: summaries, correspondence, data analysis, document review. Copying a ready block and swapping the context takes less than 1 minute.
Checking before sending
A model’s output needs checking before it goes further. The minimum checklist covers numbers, dates, names and quotes from source documents. It helps to ask the model to point to the fragment of the source each claim comes from, and then to check two or three items at random.
How often to check depends on the cost of an error. For internal correspondence, spot checks can be thinned out after a few weeks of clean results. Numbers in contracts and board reports get a full check on every version.
A separate question is what goes into the chat window at all. Personal data under GDPR, commercial terms and confidential material require a tool approved by the organization and knowledge of where the data is processed and how long it is kept.
What a better prompt won’t fix
The structure of a prompt helps only if the task can be done with the material provided. It reduces the risk that the model fills in missing data with its own guess, but does not remove it. It will not fix a badly posed question that the author cannot answer either. Nor will it replace access to systems: when the information you need lives in a database, an HR system or a document repository, the solution is an integration, for example through MCP (Model Context Protocol), a standard for connecting models to company data sources.
The starting point is simple: pick 3 tasks repeated every week, write a five-part prompt for each, test it for a few days and improve it after every failed result. We work through the team’s own tasks and prepare a prompt library and review rules as part of an AI Workshop.
Frequently asked questions
Does writing a role such as "act as an expert" improve AI answers?
The role alone does not improve accuracy on factual questions. It helps only together with context, a task, an output format and an example of a good answer.
How long should a prompt be?
Long enough that the model does not have to guess. It should include the information the task needs, because excess irrelevant data lowers answer quality.
What should you do when AI gives a bad answer?
Instead of fixing the answer by hand, add the missing instruction to the prompt, such as a word limit or a requirement to give numbers, and save the improved prompt for reuse.
Footnotes
-
Zheng et al., measurement of how a role written into the prompt affects answer accuracy: https://arxiv.org/abs/2311.10054. ↩
-
Shi et al., “Large Language Models Can Be Easily Distracted by Irrelevant Context”, ICML 2023: https://arxiv.org/abs/2302.00093. ↩
Blog content is informational and educational. It does not constitute legal or tax advice, nor individual business advisory. The scope of our services is described in the terms.
See the package: AI Workshop