A good prompt is not a spell. It is a task brief for the model. If you write “do some marketing,” the model will honestly return fog. If you provide context, a specific task, format, constraints, and quality criteria, AI starts working like a junior specialist under supervision.
A prompt is a brief, not magic
This is the second part of the series. In the first part we covered models and context: why choosing the right tool matters and why one endless chat becomes a landfill.
Now the main point: the model does not read your mind. It does not know what is in your head, what was agreed with the client, what tone your brand uses, or why “short” means 700 characters to you and not three paragraphs.
When you give a person the task “make a presentation,” a good employee asks questions. AI often does not ask; it tries to guess. A prompt should remove uncertainty before the model starts guessing.
What a working prompt needs
Context. Company, product, audience, situation, constraints, source data. Do not be afraid of length if it is structured. Models work better with specifics than with gaps.
Task. Not “analyze churn,” but “find three likely reasons for Q4 customer churn, separate facts from hypotheses, and suggest actions with a budget under 500k.”
Output format. List, email, table, plan, two-page document, presentation outline. If you do not define the format, the model will invent one.
Constraints. What not to do: do not invent numbers, do not use unverified sources, do not change legal wording, do not exceed 1,500 characters.
Quality criteria. What makes the answer good: sources are included, risks are visible, alternatives are listed, fluff is removed, facts are separated from assumptions.
Microsoft’s prompt engineering guide highlights basic techniques: start with clear instructions, break tasks down, and specify output structure.
“Start with clear instructions.” — Microsoft Learn
In plain English: the model should not have to guess what success means. Less guessing means less fluff.
“You are an expert” is not the main lever
People used to start every prompt with “you are a senior marketer with 15 years of experience.” Sometimes a role helps set tone and domain, but it does not save a weak task.
Weak version:
You are an expert. Make a promotion strategy.
Working version:
Below is a description of a B2B SaaS for CFOs. Prepare a 90-day promotion plan. Constraint: budget under $3,000, team is the founder and one marketer. Format: 5 channels, hypothesis, test cost, success metric, risk. Do not invent market numbers without a source.
The difference is not a prettier title for the model. The difference is input control.
Habits that actually help
Write the working instruction in English if the model follows it more reliably. Add Respond in Russian. at the end if needed. This is not magic, but on some models English instructions produce more predictable structure.
Ask AI to write the prompt for you. Open a separate chat: “Create a detailed prompt for task X that I will send to another model. Ask clarifying questions if data is missing.” Often this is faster than forcing the structure yourself.
Work in iterations. The first answer is a draft. A normal process sounds like: “too generic,” “add sources,” “cut it in half,” “separate facts from hypotheses,” “rewrite for a client without technical jargon.”
One chat, one task. If you wrote a contract, ad copy, and strategy in the same thread, do not be surprised when the model starts mixing contexts.
Speak instead of typing when the task is long. Voice input saves time when you need to unload context from your head: what happened, why it matters, what constraints exist.
Anthropic’s documentation notes that Claude Console includes prompt generators, templates, variables, and a prompt improver.
“The Claude Console also offers prompting tools—prompt generator, templates and variables, and prompt improver—to help you build and refine prompts quickly.” — Anthropic Claude Docs
Translation: you do not have to heroically write the perfect prompt from scratch. You can use tools and templates — but you still need to understand what you want.
How to control the result
The most dangerous mistake is accepting confident text as truth. A model can sound convincing and be wrong. For work tasks, I ask it to:
- separate facts from assumptions;
- show sources for external data;
- list risks and weak points;
- provide an alternative option;
- state where data is missing.
If the task is important — contracts, finance, legal text, medical topics, public analysis — human review is mandatory. AI prepares the draft; responsibility is not delegated.
A universal prompt template
Context:
[Who we are, product, audience, situation]
Task:
[Exactly what needs to be done]
Input data:
[Facts, documents, links, constraints]
Output format:
[List / table / email / plan / JSON]
Quality criteria:
- separate facts from assumptions
- do not invent numbers
- mention risks
- ask questions if data is missing
Respond in Russian.
This template is not perfect for every case, but it is better than 90% of “make it nice” requests.
Where to find good examples
You do not need an expensive course to start. Take a few open prompt libraries, study the structure, and adapt it: Anthropic Prompt Library, Microsoft documentation examples, and open GitHub collections.
Do not copy prompts blindly. Look for the blocks: context, format, constraints, examples, quality checks. Transfer the principle, not the text.
The short version
A prompt is controlled input into the system. The better you describe context, task, format, and constraints, the less the model hallucinates and the faster you get a useful draft. Not the final truth — but material you can work with.