If you ask the neural network, 'write a post about coffee,' you'll get a cliché like 'Coffee is more than just a drink.' Ask differently — you'll get text worth publishing. The difference isn't in the model. The difference is in the structure of the request.
Over two years, I've sifted through thousands of prompts: my own, others', from courses, from repositories. I identified seven elements that appear in every successful request. It's not a secret or magic. It's a framework on which you can build any task.
Element 1. Role
The model is an actor without a script. Until you assign a role, it plays 'average assistant.' An average assistant writes average.
Bad: 'Write a post about coffee.'
Good: 'You are a copywriter for a specialty coffee shop, 8 years of experience, writing in the style of short emotional posts.'
The role sets the vocabulary, tone, rhythm. With one sentence, you narrow the range of responses by dozens of times.
Element 2. Context
The model doesn't know who you're writing for and why. Explain:
- Who the audience is (age, interests, experience)
- Where the publication will be (Instagram, email, landing page)
- What the goal is (awareness, sales, subscription)
- What has already been said before
Context is what you keep in mind by default, but for the neural network, it's a blank screen. Spend 3 minutes — save 10 iterations of edits.
Element 3. Task
One thought in one prompt. Formulate the task in a separate paragraph so it doesn't get blurred.
Задача: написать пост в Instagram длиной 200–300 слов, который запускает обсуждение в комментариях через острый вопрос и ведёт к записи на дегустацию через био.
The task has three pillars: goal (discussion + recording), length metric, specific action. Without them, the model fills in the gaps.
Element 4. Input Data
If there's material to process — upload it. Photos, text, tables, transcripts. The more raw data, the less the neural network has to invent.
Data goes inside separators. Three quotes, tags, explicit 'INPUT DATA:' and 'END.' This saves from confusion when the model thinks part of the data is an instruction.
Element 5. Output Format
Without a team format, the neural network responds in paragraphs. Most of the time, this is not what you need.
- "Respond with a table of 3 columns"
- "Output JSON with the structure {title, hook, body, cta}"
- "Bullet list of 7 items"
- "Two options: A and B"
Especially critical if the result goes into a script, Google Docs, or a spreadsheet. A clear format saves an hour on reformatting.
Element 6. Limitations
Anti-list. What you definitely don't want.
- "Without words: innovative, unique, best"
- "No introductory phrases like 'Of course!' and 'Great question!'"
- "Don't use emojis"
- "Don't mention competitors"
- "Length — no more than 280 characters"
Limitations act as a filter on the output. The model won't consider these options, even if they come up.
Element 7. Examples
The strongest part. One good example works better than ten instructions.
Add few-shot: "Here's an example of what I like." And separately: "Here's an example of what doesn't fit." The model will pick up the style more accurately than from a description.
Пример хорошего поста: «Вчера клиент спросил: „А у вас зерно обжарено под эспрессо?" Я ответил: „Нет, у нас зерно обжарено под вкус". Повисла пауза. Через минуту он взял две пачки домой. Мы не обжариваем под оборудование. Мы обжариваем под то, чтобы зерно звучало. А как прозвучит — эспрессо или воронка — задача зерна.»
How to put it all together
[РОЛЬ] Ты — ... [КОНТЕКСТ] Аудитория ..., платформа ..., цель ... [ЗАДАЧА] Сделать ... [ВХОДНЫЕ ДАННЫЕ] """ ... """ [ФОРМАТ] ... [ОГРАНИЧЕНИЯ] Без ... [ПРИМЕР] Хорошо: ... Плохо: ...
Not bureaucracy. A map that guides the model to the desired answer. Skip a point — you miss a landmark.
Typical mistakes
Overloading in one paragraph
When you cram the role, context, and task into one chunk — the model doesn't understand what's important. Break it into sections. The structure of the prompt affects the structure of the response.
Expecting mind reading
"Well, you understand" works with people. With the model — no. Everything not written is guessed randomly.
One iteration
Even the perfect prompt rarely gives the perfect answer right away. Be prepared for 2–3 revisions. This is not a failure, but part of the process.
How to check the prompt before launching
- Read it aloud. If you stumble — the neural network will stumble too.
- Check: are all 7 elements present?
- Is there at least one example?
- Is the output format specified clearly?
- Are there prohibitions, not just permissions?
If everything is 'yes' — launch it. If not — add 30 seconds and save half an hour on edits.
FAQ
Do I need to write the prompt in English?
Claude and GPT work equally well with Russian. Write in the language in which you express yourself more accurately. The difference in quality is minimal.
How long should the prompt be?
As long as necessary for the 7 elements. Working requests are often 300–800 words — and that's normal.
Why does one prompt give different results?
Models are probabilistic. Strengthen the role, format, examples. This narrows the corridor.
What to do if the neural network goes off track?
Add constraints: banned words, length, format. Specify the role and audience. Show examples of good and bad responses.
Can you learn over the weekend?
The basic structure is yes. The feel for the model comes after 200–300 prompts. The skill is like interviewing.
What's next
The formula is the skeleton. Next, you need practice on real tasks: copywriting, sales, analytics, strategy. I have compiled a system of 50+ prompts with explanations of what works and why in each case.