How to write prompts for ChatGPT in 2026?

Use a structure of 6 elements: role, context, task, output format, few-shot examples, restrictions. On my projects, this approach gives results on the first try in 70% of cases versus 20% without a structure.

How is chain-of-thought different from few-shot?

Chain-of-thought - a request to the AI to reason step by step (“think step by step”). Few-shot - give examples of the correct answer to the task. The first is better for analytics and calculations, the second is better for copywriting and creatives. The main thing: CoT is not needed when the task is simple, few-shot does not work when there are more than 5 examples.

Which AI is better for writing marketing prompts?

Claude - for data analysis, long texts, strategies. ChatGPT - for creativity, headlines, short texts. Perplexity - for research and fact checking. Gemini - for multimodal tasks and working with tables. On my 6 projects, Claude closes 50% of tasks, ChatGPT - 30%, Perplexity - 15%, Gemini - 5%.

How much does it cost to use AI for marketing per month?

AI stack budget for a marketer in 2026: Claude Pro $20, ChatGPT Plus $20, Perplexity Pro $20, Claude Code $20 - total 80-150 $/month. Saves 12-18 hours per week. On my projects, the ROI of the AI stack is 5-8x due to the reduction in time for copywriting, analysis, and creatives.

What prompts does a marketer need every day?

Of my 7 working templates: competitive teardown, semantic core, content plan, cohort analytics, A/B hypotheses, brief analysis and creative iterations. Read more in the article “7 Prompts for a Marketer.”

What is a system prompt and why is it needed?

A system prompt is an instruction that is executed before each request. For a marketer: “You are a performance marketer with 8 years of experience, working in the Russian market. Always give an answer with numbers, without general phrases. Use “Christmas trees” for quotation marks. Do not write “in the era”, “plays a key role”, “in the modern world”. The system prompt sets the tone of voice and the quality of the base from which each response is based.

How to check that a prompt is written well?

Three criteria: 1) the answer is close to the right one the first time (no need to rewrite 3-4 times), 2) with the same prompt, AI produces consistently similar results, 3) changing the model (Claude instead of ChatGPT) with the same prompt gives comparable quality. If any of this is wrong, the prompt needs to be modified.

What are the most common errors in prompts?

Typical: 1) no role - the answer is generalized, 2) no context - AI does not understand the task, 3) no examples - the result does not fall into the tone of voice, 4) the prompt is too long - AI loses focus on the 3-4th paragraph, 5) no restrictions - AI does extra work.

All articles

AIMay 26, 202610 min

Prompt engineering for marketers 2026: 6 elements + 8 ready-made templates

Step-by-step guide to prompt engineering for marketing: 6 elements of a good prompt, 3 techniques (few-shot, chain-of-thought, system prompts), 8 ready-made templates for different tasks. With real examples of prompts for Claude and GPT.

Most of the prompts that my colleagues show me are “write a post about AI in marketing.” AI answers in general phrases because the question is general. I’m not looking at “how to write a prompt,” but specifically: what 6 elements a working prompt consists of, how the techniques differ, and what 8 templates I’ve had in Notion for two years now.

I have been working with AI in marketing since the summer of 2024. On 6 current projects, Claude, ChatGPT and Perplexity cover 70% of recurring tasks. The numbers in the article are from my performance campaigns with a budget of 1M ₽/month. Time savings: 12-18 hours per week, cost of the AI stack is $80-150/month.

A bad prompt is not a “short prompt”. This is a prompt in which the AI guesses what you wanted to say instead of knowing.

1. Six elements of a work prompt

Over the course of two years, I tried a bunch of formats. One structure of 6 blocks worked. It is not necessary to write all 6 every time - for simple tasks 3-4 are enough. But when a prompt doesn’t produce a result, the problem is almost always in one of the missing elements.

Role — who does AI work. Not “you’re a marketer,” but “you’re a performance marketer, you’ve been working in the Russian market since 2017, you specialize in Yandex Direct and VK Ads, you manage budgets starting from 1M ₽/month.” A specific role changes quality by 2-3 times.
Context — introductory: product, niche, audience, current metrics, budget, what has already been tried. Without context, AI makes a “hospital average.”
Problem - what exactly needs to be done. Not “analyze competitors”, but “find 5 competitors in the EdTech niche, compare their advertising strategies according to 4 parameters: channels, offer, CPL, audience.”
Output Format — what the response should look like: table, list, JSON, text. Indicating the structure: “three headline options”, “100 word summary”.
Examples (few-shot) — 1-3 examples of the correct answer. Few-shot reduces the number of iterations by 3-5 times.
Limitations - what not to do: do not write general phrases, do not use “synergy”, do not mention competitors by name.

2. Few-shot, chain-of-thought and system prompts

There are three main techniques in prompt engineering 2026: few-shot, chain-of-thought and system prompts. Each one has its own task.

Few-shot. You give the AI 1-3 examples of what the answer should look like. Works for copywriting, creating headlines, product descriptions. Examples must be real - not fictitious. I store 20+ examples of good tone of voice in Notion and substitute them into task prompts.

Chain-of-thought. You ask AI to reason step by step. Not “calculate ROAS for the quarter,” but “first divide the income by month, then calculate the costs for each source, then divide it.” Works for analytics, calculations, building strategies. The main thing is not to use CoT for simple tasks: AI overcomplicates the answer.

System prompt. An instruction that is executed before each request. I use one system prompt for all tasks: “You are a performance marketer with 8 years of experience in the Russian market. Answer with numbers. Use Christmas trees. Do not write “in the era”, “plays a key role”, “in the modern world”. In Claude this is set in the project settings, in ChatGPT - in Custom Instructions.

3. Which model to use for which task?

There is no universal AI. My distribution on projects is as follows: Claude - 50% of tasks, ChatGPT - 30%, Perplexity - 15%, Gemini - 5%. And here's why.

Problem	Which model is better	Why	Save time
Strategy, brief, longread	Claude Sonnet 4.6	Holds context of 200K tokens, structural analysis	3-4 hours → 40 min
Headlines, creatives, storytelling	ChatGPT 4o / 4.7	Better creativity, more freedom in wording	2 hours → 20 min
Research, fact-checking, data	Perplexity Pro	Searches in real time, provides sources	3 hours → 30 min
Multimodality, tables, graphs	Gemini 2.5 Pro	Better work with images and files	1-2 hours → 15 min
Code, n8n scripts, automation	Claude Code	Writes production code, works in the terminal	1 day → 3 hours

Read more about the difference between models in my comparison ChatGPT vs Claude vs Gemini. There is a table with benchmarks for 10 tasks.

4. Template 1-2: competitor analysis and semantic core

Pattern 1: teardown competitors. I enter into Perplexity: “Find 5 competitors [niche]. For each: promotion channels, offers, audience, CPL (if available in open sources), strengths and weaknesses.” Perplexity returns a structured answer with sources in 5 minutes. I used to do this by hand - 3-4 hours.

Template 2: semantic core. “You are an SEO specialist for the Russian market. Product: [description]. Nisha: [niche]. Create a semantic core: 50 queries grouped by intent. For each – frequency (high/medium/low without exact numbers).” Claude produces a clustered kernel in 10 minutes. I'm just adjusting the grouping.

35 more tasks with prompts - in the article ChatGPT for marketers: 30 tasks. There I analyze a complete set of templates for each direction.

5. Template 3-5: content plan, analytics and hypotheses

Template 3: monthly content plan. “Product: [description]. Target audience: [description]. Goal: [leads/reach/brand]. We need a content plan for 30 days, broken down by categories: expert content, selling, engaging, reputational. For each post - title, format, purpose, CTA." Claude provides the structure, I provide the expertise.

Pattern 4: Cohort analytics. “I have data: [table]. We need to: calculate the retention of cohorts by week, determine the point of outflow, propose 3 hypotheses why retention drops after the 4th week.” Chain-of-thought + Claude provides analysis in 15 minutes, which previously took 2-3 hours in Excel.

Pattern 5: A/B hypotheses. “Landing: [description]. Current conversion: [X]%. Traffic: [Y] visits/month. Propose 5 A/B hypotheses with ICE priority. For each – what we change, how we measure it, the expected effect.” Claude produces structured hypotheses, I select the top 2 for the test.

6. Template 6-8: creatives, briefs and iterations

Pattern 6: Creative Iteration. “You are the creative director of a performance agency. Product: [description]. Audience: [description]. Channel: [VK Ads / TG Ads / Yandex Direct]. Write 10 variations of headlines and 10 variations of ad texts. Use the format: problem → solution → result → CTA. Each option is up to 90 characters.” Claude gives 10 options in 2 minutes. Of these, 2-3 are workers. This is fine.

Template 7: brief analysis. I give Claude a brief from a client: “Analyze this brief. Find: 1) what is missing to start, 2) what questions to ask the client, 3) what risks you see. The answer is a 10-point checklist.” Saves 30-40 minutes on each new client.

Pattern 8: Content Iteration. “Here is the text: [text]. Rewrite it: 1) keep it to 100 words, 2) add numbers from the context, 3) remove common phrases, 4) make the tone of voice like [link to example].” I use this template 10-15 times a week.

More templates in the article 7 prompts for a marketer that I use every week. There, each template is analyzed with a real example of the output.

7. Five common mistakes in prompts

1. No role. “Write a post” - AI writes like AI. “You are a CMO with 8 years of experience in performance marketing in the Russian Federation” - AI writes as a CMO. The difference in quality is 2-3 times.

2. No examples. AI focuses on the average temperature in the hospital. If you don’t show an example of the desired style, you’ll get template text. Few-shot is not an option, but a mandatory element.

3. The prompt is too long. After 500-700 words the AI starts to lose focus. If the prompt is for 3 screens, split it into several consecutive ones. I have a rule: one prompt = one task.

4. No restrictions. AI does extra work. Write “don’t offer something that no longer works: Instagram in the Russian Federation, Google Ads without a VPN, email without a database.” This saves 2-3 iterations per prompt.

5. The same model for everything. ChatGPT is not the best for analyzing long texts. Claude is not the best for short creatives. Perplexity is not a research tool, but a search engine. I keep three tabs open and switch to task.

8. Conclusion

Prompt engineering in 2026 is not magic, but structure. Six elements, three techniques, five models for different tasks. 8 templates from my practice cover 80% of a marketer’s repetitive tasks.

The main thing I’ve learned in two years is that a prompt doesn’t have to be perfect the first time. 70% of quality is enough - the rest is achieved through iterations. But without structure there are 5 times more iterations.

If you want to analyze your specific situation, write to Telegram or through form. Starting consultation - 0 ₽.