7 prompts for a marketer that I use every week

Once every two weeks I receive the same message: “drop off your folder with prompts.” There is no folder. There are seven templates in Notion that I've been returning to 3-5 times a week since the summer of 2024. This is not the “top 50 prompts for a marketer” that any neural network can generate for itself in half an hour. These are seven pieces that passed the filter “works on combat projects with a performance budget of 1M ₽ per month.” All seven are below, without abbreviations.

Disclaimer at the entrance: “prompt engineering” is half a marketing word, half a discipline. Marketing - because in most courses about it they sell you what Claude already does by default. Discipline - because **the difference between a “help with creativity” question and a work prompt is the difference between an hour lost and a week saved**. In my cases, this difference is already calculated in real money, and I stopped arguing about “a tool or a replacement for a person.” Let's get down to business.

Prompt is not magic. This is a pattern that repeats between tasks. If you don’t have 5-10 repeating patterns in your work, you’re not using AI, you’re communicating with it.

1. Competitive teardown in 30 minutes instead of 2 days

When: before entering a new project and once a quarter for current ones. Until 2024, I did the following analysis manually: I opened 5–7 competitors’ websites, looked at social networks, read reviews, and wrote them down on a tablet. It took 2 working days. Now - 30 minutes per prompt + 30 minutes for verification in sources.

What it gives: positioning of a competitor in three axes (product, audience, price), a list of 3–5 message lines that it covers, and a list of 3 message lines that it not covers. The last thing is the most valuable: a hole in a competitor’s communication is your first candidate for a USP.

The prompt template I have right now in Notion is:

You are a competitive analyst working with a marketer. I give you data about the competitor: website, product description, price list, 5 latest advertising creatives, 10 latest posts on social networks.

Your task is not to describe your competitor, but to draw out his communication model.

Do exactly three things:

1) Pos triangle: positioning in three axes - product (features vs benefits), audience (to whom they are selling according to the description), price (premium / mid / budget). For each axis - a quote from their own material on which you base your conclusion.

2) 3–5 message lines, which they actively use. Each one has a proof: where it was taken from.

3) 3 message lines that they DO NOT cover, but which can legitimately be used against them (that is, the audience asks about this in reviews / comments / searches).

There is no “is the market leader”. Just specifics.

The output format is Markdown with headers, without filler.

Where it merges: if the competitor is closed (b2b enterprise without public communication) - there is little data, conclusions become strained. Then I download employee reviews from Habr Career / Glassdoor - this often restores the picture.

Real metric: on the MM AI Trading project (fintech, UAE), such a teardown on 4 competitors helped to find the position hole “trading for AI-native audience” - none of these four occupied it. I built a launch around this hole.

2. Rewrite someone else's text into your voice

When: client brief, landing page, series of letters. Any situation where you need to take a wording from someone else’s head and put it into your own voice field, without re-inventing it.

The main trick here is not the prompt itself, but context. I have a folder in my site repository /content/voice/ with 11 files for 50K characters: voice coordinates according to the NN/G matrix, five rhetorical pillars, dictionary, rhythm, list of AI-tells. Before each rewrite task, I load this entire folder into the Claude context - and only then submit the source.

Context: [attaching the entire folder /content/voice/]

Next is the text that needs to be rewritten into this voice. The text is NOT mine, it was sent to me by a client/partner/colleague, and it is written in a neutral corporate intonation.

[insert source]

Task:
- Save all facts and figures from the source without loss
- Convert intonation to coordinates Casual 3 / Funny 6 / Irreverent 7 / Matter-of-fact 2
- Apply at least 4 of 5 pillars (antithesis, number, herringbones on a stamp, equation, insider term)
- DO NOT add your own facts that are not in the source code
- DO NOT add “worth noting”, “in the modern world”, “is” - see file 06-ai-tells.md

Give two options: hard (maximum voice) and soft (50/50 between the source and voice).

Where it merges: when the client is premium B2B (private banking, legal consulting), and the text is required to have a deliberately boring, formal tone. Then my voice works against the task. In such cases I do reverse transfer - I feed the client’s Claude voice and rewrite my drafts there.

Metric: on this article, the first paragraph went through the rewrite template three times - because Claude first inserted five stamps from 06-ai-tells, and only the third pass cleared everything out. Without a template, I get the same result by hand in 20–25 minutes.

3. 30 headlines in three logics per request

When: launching a new campaign in VK Ads, Yandex Direct or Telegram Ads. Any situation where you need 10–30 headline options with different rhetoric for an A/B test.

Standard error: “give 30 headlines for X ad.” The end result is 30 variations of the same thought, rearranged like solitaire. The usefulness is zero, there is no variety for the test.

Working version - division into three rhetorics in one request:

Product Context: [short description, ICP, 3 main benefits, 3 objections]

Generate 30 headlines for VK Ads (length - up to 33 characters with spaces), divided into three groups of 10:

GROUP A - logic. The headline is based on a number or comparison. No emotions. Example of a pattern: “CAC decreased by 40% in a quarter.”

GROUP B - conflict. The headline puts the reader in a position of choice, debate, or provocation. Example of a pattern: "Your performance is not working. Here's why."

GROUP C - benefit in the moment. The headline promises measurable action in the short term. Example of a pattern: "30 minutes of analysis. Free."

LIMITATIONS:
- Do not use words: unique, effective, innovative, advanced, best
- Do not use exclamation marks
- Don't use emojis
- Each headline is self-sufficient, without reference to others

After 30 headlines - a short summary: which group works better in which niches in my experience (based on publicly available data on performance benchmarks).

Where it bleeds: AI has a poor sense of cultural context and puns. If the campaign requires a play on words or a local reference (such as a reference to gaming slang or the vibe of the area), it misses the mark. I still write this part with my hands.

The metric for checking headlines is CTR with comparable audiences and a sufficient volume of impressions. Then compare leads and sales: the winner in clicks is not necessarily the winner in business results. For NEMIFIST, there is no report on such an A/B test in the available materials.

4. Hypotheses for the reasons for churn based on the cohort table

When: retention falls, LTV does not grow, or the client complains “I don’t know why they are leaving.” Before AI, it was a week of work as an analyst and another week of interviews with those leaving. Now - an hour for the prompt, two hours for testing hypotheses, two days for final interviews with the most likely candidates for the cause.

The main trick is to never ask Claude about churn in general terms. First you feed the cohort table raw data - step by step, so that he builds a vision of the dynamics, and only then ask for hypotheses.

Step 1. I submit the retention cohort table by month (CSV, 12+ cohorts). Data only, no questions asked.

Step 2. I ask: “Describe what you see in this table. Where are the anomalies, where are the patterns. Don’t interpret - just describe.” Claude responds with 200 words of observation.

Step 3. Provide the context of the product: what it is, what ICP, pricing model, last 3 product changes with dates.

Step 4. Only now - the main request:

"Taking into account the table and context, formulate 5-7 working hypotheses for the reasons for the accelerating churn. Each hypothesis should:
1) Have a reference to a specific row/cohort of the table as proof
2) Be verified in 1 day (not “the product needs to be remade”)
3) Have a proposed verification method (cohort segment, exit-survey, behavioral analysis)

Rank by expected explanatory power. Weak hypotheses - weed them out, don't stretch them to 7."

Where it merges: in small volumes. Less than 100 users in a cohort - Claude begins to hallucinate patterns. In such cases, I return to manual work.

Metric: on one of the e-com projects in 2024, this prompt produced 7 hypotheses in an hour, of which I rejected 4 immediately, 2 were checked through a cohort sample, and one turned out to be correct - a product change in April broke onboarding for returning users. Corrected in one sprint, retention was restored in 6 weeks.

5. Transcript of discovery call → strategy structure

When: After the first 30-60 minute call with a potential client. Previously, after such a call, I would sit down for half a day and collect a summary + first draft strategy. Now - 15 minutes.

I take the call transcript from Granola or Otter - they both provide clean markdown. Next:

Context: I am a freelance marketer, I conducted a discovery call with a potential client. The transcript is below.

[full transcript]

Do three things:

1) ICP map: who the client described as a buyer of the product (one line), who he really wants to attract (often different, read between the lines), and where the gap is.

2) Top 3 real pains of the client (what he repeated several times, or spoke with irritation, or quickly changed the topic). I know that declared pain and real pain often do not coincide - find the real one.

3) 90 Day Plan Framework:
   - Month 1: one hypothesis tested over 4 weeks
   - Month 2: Deepening what worked in M1 or pivoting
   - Month 3: scaling

   Without bullet lists “do SMM, performance, content, analytics.” Specific steps with outcome metrics for each month.

And one more thing: 3 red flags in this transcript - that working with this client will be difficult. If there are no red flags, say so.

The main thing in this prompt is the last point. Claude really knows how to pull red flags out of a transcript because he has no vested interest in taking on the project. I have. And this is the first time that an AI draft warned me about a client whose “flexible budget” in the first conversation turned out to be a trap for concession-overload.

Where it drains: on short calls (15 minutes). The transcript is too thin, the conclusions are strained. Calls last 45+ minutes - it works perfectly.

6. JTBD unpacking of user interviews

When: After a series of 5-10 interviews with current or lapsed clients. JTBD (jobs-to-be-done) is a methodology in which you don’t ask “what do you like,” but reconstruct the moment “I switched / bought / left”. This provides real triggers, not desired responses.

Each interview I have is a markdown file of 1500–2500 words. Serve them to Claude one at a time - slowly. Serve in a bunch - it blurs the context. Solution:

I have 8 interview transcripts, one per message. I'll upload them one by one.

After EVERY interview, you do exactly one thing: pull out trigger moments. A trigger is a specific event or phrase that led to a change in behavior (bought, abandoned, switched, found, changed my mind).

Format for each interview:

[Interview ID]
- Trigger 1: [quote] → [job type]
- Trigger 2: [quote] → [job type]
- ...

DO NOT add up. DO NOT generalize. Just triggers with quotes.

After the 8th interview - final step:
- Cluster all triggers into 3-5 groups according to job type (functional / emotional / social)
- For each group - what is the capacity in the sample (how many interviews contain it)
- Final list: 3–5 jobs, ranked by frequency

I will use this list as a messaging map for the landing page.

Where it merges: if the interview was conducted poorly - without specific time references, without questions about “tell me about the last time when...” - Claude still tries to get triggers where there are none, and draws conclusions. The quality of the output is equal to the quality of the questions at the input.

Metric: on this prompt, I usually collect a messaging map in 90 minutes instead of the usual two days of manual analysis.

7. Segmentation from unit-economics, not from demographics

The most expensive lesson in my career is described in a separate article - case of failure by −20% to the quarterly plan. Segmented by demographics, should have been segmented by behavior. I collected this prompt exactly after that case, like “never again.”

When: CAC grows, ROAS falls, you need to reallocate the budget between segments. Any situation in which there is a temptation to open Ya.Metrika and cut the audience into “18-24 / 25-34 / 35-44”. Don't do that. Do this:

Context: project [description]. ICP is broad; a single performance strategy does not produce results.

I have data:
1) AOV for the last 12 months (distribution, not average)
2) Frequency of purchases by cohorts
3) Time-to-second-purchase (if available)
4) Recruitment channel by segment
5) Retention curve by segments

[insert CSV or table]

The goal is not to cut into segments based on demographics. Cut them up so that within each segment the unit-economics are homogeneous.

Specifically:
- Divide the audience into 3-5 behavioral segments
- For each, calculate: average CAC, average LTV, payback period
- Rank by LTV/CAC ratio
- Show where the money is “drowning” - the segment into which the budget flows, but the unit economy is negative
- Show where the money is “underinvested” - a segment with high LTV/CAC, but a small share of the budget

Final: recommendation for budget redistribution between segments for the next quarter. No “increase coverage”. Specific percentages.

Where it merges: if there is little data (less than 6 months of history and less than 500 purchases) - Claude pulls patterns on noise. In such cases, I honestly tell the client that segmentation requires another 3 months of data, but for now we are working on a hypothesis.

What I stopped using

Over the 3 years of working with the “marketer + AI” combination, several patterns that I relied on at the beginning died.

• Prompt chains for 10 steps. Sounds cool, doesn't work well. In the third step, Claude loses context and begins to hallucinate. Now I keep prompts as atomic as possible - one prompt = one task.
• “Play the role of CMO of Coca-Cola.” Role-prompts worked in 2023 with GPT-3.5. With Claude 4.6 they give formulaic answers - the model plays "CMO" like a caricature.
• Long system prompts with ten rules. Claude ignores eight of the ten rules. Better - three strict rules and verification through follow-up.
• “Be critical” / “Critique my idea.” The model is criticized superficially - it is optimized to “help”, not to “break”. Real criticism comes only from living colleagues.

What do I have left outside the prompts?

All seven templates above work because I keep the context in mind. The prompt is half the task. The second half is understanding what is considered an answer and what is garbage that Claude generated so as not to leave me without a way out.

This means no “just run the prompt and publish”. Every article, every creative, every strategy after AI goes through my editing - where I throw out tells, check the numbers, and often completely rewrite one of the paragraphs because Claude guessed the topic, but didn’t guess the intonation. This takes ~30% of the time that I would spend working from scratch.

If you were looking for a “folder with prompts that will solve marketing for you,” there is none. Not with me, and nowhere. The seven patterns above are seven levers that work in my hands, because I know what counts as an answer. In the wrong hands, they will give less - exactly the difference that separates the tool from the skill.

If you want me to analyze your work flow and put together the same library for your tasks, write in the form of contacts, I allocate 30 minutes for free for discovery calls. If you want to see how this stack works in production, take a look case “One marketer = team of 5 roles”, all the arithmetic is there.

Related materials: unit economics calculator, content plan template, a complete guide to performance marketing.