# Prompts for Marketers: Templates, Tests, and Compliance 2026

- URL: https://vladimirnovozhilov.com/en/blog/prompts-for-marketers-templates-tests-and-compliance-2026/
- Author: Владимир Новожилов
- Published: 2026-10-03
- Category: AI & vibe coding

> This single guide contains a library of prompts for marketers, an algorithm for creating them, A/B testing methodology, and a compliance checklist. Includes before/after examples, model recommendations for Russian tasks, and a ready prompt‑book in CSV/JSON. Suitable for B2B, IT and services.

Businesslike: prompts for a marketer are not a "magic phrase" but a proper brief for an intelligent executor. AI quality depends on the task setup: who it is (role), what it knows (context), what it must do (task), how to report (format) and what to avoid (constraints). This article contains a practical library of templates for content, SMM, ads, email and analytics; a working algorithm for writing prompts; A/B tests and metrics; model selection for the Russian context; and data security. All examples and recommendations are current as of 2026.

## Who this material is for and what it contains

Short on benefits.

- Who: marketers in B2B, IT and services, editors, media planners, performance managers, team leads and CMOs.
- What you'll get: a library of ready prompt templates, before/after examples for each task category, a validation checklist, A/B testing methodology with KPIs (CTR, CR/CVR, CPA, CAC, LTV), model recommendations (ChatGPT/GPT‑6.x, Claude Sonnet/Opus, YandexGPT, GigaChat), safe prompts without personal data, and a prompt‑book in CSV/JSON.
- When to use: when building a content plan, ads, emails, funnels, JTBD maps, competitor analyses, or reports to leadership.

## What a prompt is: structure and quality criteria

Practical definition: a prompt is a formalized brief to a model with role, context, task, response format and constraints.

Structure of a strong prompt (remember this formula):

- Role. "You are a B2B editor / funnel analyst / media planner" — sets style and depth.
- Context. Industry, audience, JTBD (what job the product is hired to do), funnel stage, SEO intent.
- Task. What exactly to produce and why: "generate 10 A/B hypotheses with priorities."
- Format. Table / bullet points / draft text / SQL / brief with fields.
- Constraints and quality criteria. Tone, length, factuality, "no PD, no links", "examples from Russia 2024–2026", "3 alternatives".
- Data. Source materials: theses, exports, campaign parameters, existing text — better to include as separate blocks.
- Success metric. Which KPI will be used to evaluate the result: CTR, CR, CPA, CPL, NPS, speed of delivery.

Definition of Done (acceptance criteria):

- Match role and brand tone. 9/10 phrases sound like a real expert, no clichés.
- Hit the intent (SEO intent / funnel stage). No vague answers.
- Format as specified: tables, length, CTA, structure.
- Verifiability of facts/figures: no hallucinations, questionable data marked.
- Adaptation to Russian audience and 2026 local realities.

Template matrix (copy as a base):

```
Role: [who you are + who you write for]
Context: [industry/market], Audience [segment, JTBD], funnel stage [TOFU/MOFU/BOFU], SEO‑intent [info/commercial], product/service [one line]
Task: [what to do + goal]
Format: [output type, fields/headings, length]
Constraints: [tone, no PD, no links, verify facts]
Source data: [paste theses/tables]
Quality criteria: [3–5 items]
Success metric: [KPI and threshold]
```

## How to write a working prompt: step‑by‑step algorithm (with examples)

1) Clarify the goal and metric. Example: "Reduce CPA by 15% without volume loss."
2) Gather context. Market, audience segment, JTBD, competitors, current creatives.
3) Choose the model role and output format.
4) Set constraints and quality criteria.
5) Split complex tasks into stages: research → hypotheses → draft → editing.
6) Attach source data as a separate block.
7) Run a mini‑iteration: request → edit prompt → final version. Save the prompt in the library.

Before/after example for an ad:

Before (lazy prompt):

```
Write ad text for contextual advertising for CRM implementation services.
```

After (working prompt):

```
You are a B2B performance editor.
Context: service — CRM implementation for mid‑sized companies (50–250 employees), industries: distribution, service. Audience — heads of sales and operations directors, JTBD: "transparent funnel and predictable forecasts". BOFU.
Task: 5 ad variants for contextual advertising, different mechanics (social proof, specificity, deadline, economic effect, minimized risk).
Format: table [Mechanic | Headline ≤ 30 | Text ≤ 90 | CTA | Key metric for A/B].
Constraints: no fluff or buzzwords, numbers only if verifiable, Russian language.
Quality criteria: clear benefit, concrete triggers, BOFU fit, 1 USP per variant.
Success metric: +10% CTR vs benchmark, preserve CVR.
```

This kind of setup usually reaches KPIs faster and is easier to validate.

## Prompt library by task (content, SMM, ads, email, analytics, performance)

Below — ready templates. For each block I give a before/after example. Copy and adapt to your context. Where relevant — I note recommended model and settings.

### Content prompt examples (planning, draft, rework)

1) Monthly content plan (B2B, TOFU/MOFU)

```
Role: content strategist B2B in IT services.
Context: product — IT infrastructure audit; Audience — CIOs/IT directors of mid‑sized enterprises; JTBD: "reduce IT risks and CAPEX". TOFU/MOFU.
Task: 4‑week content plan for blog and LinkedIn/Telegram: 8 articles, 12 posts, 4 short videos.
Format: table [Week | Channel | Topic | Format | Goal | CTA | Metric (CTR/ER/leads)].
Constraints: topics without hype, practical value, examples from RF 2024–2026.
Criteria: cover 3 clusters of search intents, 30% content reuse.
Model: Claude Sonnet or GPT‑6.x, T=0.4.
```

Before/after:

- Before: "Make a month of post plans." 
- After: clear audience, JTBD, channels, table format and metrics.

2) Draft SEO article for informational intent

```
You are a B2B editor.
Context: article "How to choose a SIEM for mid‑market", informational intent, cluster: security/risk reduction; Audience — InfoSec leads.
Task: outline H2/H3 + draft 1800–2200 words; highlight selection criteria, a comparison table of 5 solutions, a checklist.
Format: Markdown: TOC, H2/H3, table [Criterion | Why it matters | How to check], "In short" block.
Constraints: no promotional inserts, no links, examples for RF.
Criteria: natural term density, readability Flesch‑Kincaid for RU — average/above.
Model: GPT‑6.x or Claude Opus, T=0.3.
```

3) Rewriting text into brand voice

```
Role: brand editor.
Context: tone — businesslike, friendly, no slang; brand voice is concise and factual.
Task: edit the text below into the specified voice; reduce by 15%, increase specificity.
Format: version v1 (rewritten) + v2 (shorter, 500–700 chars) + list of replaced clichés.
Constraints: do not invent facts, keep meaning.
Source:
---
[INSERT TEXT]
---
Model: YandexGPT or GigaChat, T=0.3.
```

4) Case study "before — work — result"

```
Role: case editor.
Context: B2B analytics implementation service.
Task: compose a case study with: Client → Task/JTBD → What was done (3–5 steps) → Result (figures CTR/CR/CPA/LTV) → Business impact.
Format: 900–1200 words, "Before/After" table with 3 metrics.
Constraints: no client name, replace tags with [CLIENT], [SECTOR], amounts as ranges.
Criteria: measurability, honest baseline.
```

5) FAQ block for a service page

```
You are a product marketer.
Context: BI implementation service, audience — operations directors.
Task: 8 Q&A items: timelines, effects, risks, support, integrations.
Format: markdown list; answers 2–4 sentences, no links.
Constraints: plain language, no PD.
Criteria: address 80% of common objections.
```

6) 15 headlines + 10 first paragraphs

```
Role: headline editor.
Context: article about reducing CPA in performance.
Task: 15 H1s with different mechanics (numbers, question, benefit, "against expectations", case), 10 first paragraphs (100–140 words each).
Format: two‑part table.
Constraints: no clickbait.
Criteria: 3 headlines ready for A/B.
```

### Socials and Telegram prompt examples

1) Weekly plan for Telegram + VK

```
Role: SMM editor.
Context: B2B audience, high average check, tone — expert, no memes.
Task: schedule of 7 posts: educational, case, short insight, poll, video, FAQ, offer.
Format: table [Day | Format | Topic | Opening hook | CTA].
Criteria: 3 hooks with different mechanics.
Model: YandexGPT/GigaChat, T=0.5.
```

Before/after:

- Before: "Generate posts for a week in Telegram."
- After: clear format, tone and hook mechanics.

2) A post that doesn't feel "robotic"

```
You are a Telegram SMM editor.
Context: topic — "how we reduced CPA by 22% in 4 weeks", tone — lively, no bureaucratic language.
Task: post 1200–1600 chars, PAS structure (problem‑agitate‑solve), 1 "before/after" table, 1 CTA.
Constraints: numbers are from the block below, no guesses.
Data:
— CTR: 1.2% → 1.35%
— CVR: 3.0% → 3.4%
— CPA: 4300 → 3350
```

3) A 5‑post warming series

```
Role: SMM strategist.
Context: launching a pilot service offer "funnel audit in 72 hours".
Task: 5 posts: pain, mistake, mini‑case, method, offer.
Format: table [Post | Goal | Message | Proof | CTA].
Criteria: at least 1 social proof per post.
```

4) Replies to comments/questions

```
Role: SMM moderator.
Context: frequent question "expensive/long/complex".
Task: 6 ready replies: calm, with numbers, empathetic, humorous, redirect to DM, refusal.
Format: list by tone.
Constraints: no negative language, no promises on timing if no slots.
```

5) Visual and cover ideas

```
You are a creative editor.
Context: topic — "A/B testing mistakes in marketing".
Task: 10 cover concepts (text + visual metaphor), 3 for 9:16 video.
Format: table [Format | Idea | Cover text ≤ 6 words | Why it will work].
```

### Advertising prompt examples (search, targeting, copy)

1) 10 concepts for targeting creatives

```
Role: performance creative.
Context: CRM implementation audit; Audience — ROP/CEO in SMB; goal — lead form.
Task: 10 concepts, each — headline, key visual, offer, social proof, risk‑reverse.
Format: table.
Criteria: 4 concepts low risk, 4 medium, 2 bold.
Model: GPT‑6.x, T=0.7 (creative).
```

2) Search ads (3 mechanics)

```
You are a contextual ads editor.
Context: queries "внедрение crm", "настроить crm под ключ"; region — RF.
Task: 6 headlines (≤ 30) and 6 descriptions (≤ 90) across 3 mechanics: benefit, deadline, proof.
Format: table [Mechanic | Headline | Description | CTA].
Criteria: keywords used naturally, no keyword stuffing.
Model: YandexGPT, T=0.4.
```

3) UTM table for a campaign

```
Role: media planner.
Context: campaign "CRM‑audit", sources — Search/RSYA/VK.
Task: compile UTM parameters and naming for campaigns/ads.
Format: table [Source | Campaign | utm_source | utm_medium | utm_campaign | utm_content | utm_term | Example value].
Constraints: unified naming, no Cyrillic in UTM.
```

4) Landing A/B hypotheses

```
Role: conversion analyst.
Context: landing for CRM implementation, current form CR — 2.9%.
Task: 10 hypotheses with ICE scoring.
Format: table [Hypothesis | Insight/Data | Change | Expected effect | ICE].
Criteria: at least 3 hypotheses about social proof.
```

5) 30–45s video script (ads)

```
You are a performance scriptwriter.
Context: offer — "Funnel audit in 72 hours"; Audience — SMB owners.
Task: 30–45s script using AIDA, voiceover text, shot notes.
Format: table [Seconds | Shot | Text | Subtitles | CTA].
```

Before/after (ads):

- Before: "Write an ad for RSYA." 
- After: "Role: contextual editor; context: BOFU, segment, mechanics; table format; quality criteria; CTR metric." 

### Email and CRM prompt examples

1) Welcome sequence of 5 emails

```
Role: email editor B2B.
Context: lead downloaded a CRM selection guide; segment — mid‑market; goal — demo.
Task: 5 emails: value → case → approach comparison → objections → offer.
Format: for each email — subject (≤ 55), preheader (≤ 80), body (800–1200), CTA.
Criteria: 2 subject variants for A/B.
```

2) Re‑engagement of cold contacts

```
You are a CRM marketer.
Context: base hasn't opened emails for 90+ days.
Task: 3 emails with soft offers and a "choose frequency" option.
Format: subjects, preheaders, 3 CTAs (content/diagnostics/unsubscribe).
Constraints: no pressure.
Metric: +20% Open Rate vs benchmark.
```

3) Subjects and preheaders: 10×2

```
Role: editor.
Context: BOFU email offering demo.
Task: 10 subjects + for each 2 preheaders with different mechanics (deadline/fact/question).
Criteria: avoid "free" in subject; length ≤ 55.
```

4) Behavioral segmentation of the base

```
Role: CRM analyst.
Context: events — open, click, visit, lead.
Task: RFM‑light segmentation scheme + trigger rules.
Format: table [Segment | Rule | Message | KPI].
```

5) NPS survey (B2B service)

```
You are a researcher.
Context: clients 3+ months.
Task: NPS form + 3 open JTBD questions.
Format: email text + form questions.
```

Before/after (emails):

- Before: "Write a demo offer email." 
- After: a sequence, varied subject mechanics, segmentation and KPI rules.

### Analytics and competitor analysis prompt examples

1) Competitor analysis by site/content

```
Role: competitive analyst.
Context: service — CRM implementation, RF market; 3 competitors: [A], [B], [C].
Task: table of positions and content gaps.
Format: [Competitor | USP | Topics | Formats | CTA | Gaps vs us | Article/case ideas].
Constraints: no speculation, only observable site elements.
Model: Claude/ GPT‑6.x.
```

2) Audience map: pains/jobs/selection criteria

```
You are a JTBD researcher.
Context: Audience — heads of sales in SMB.
Task: JTBD table: [Situation | Job | Barriers | Choice criteria | Risks | Triggers].
Format: table + 5 client‑voice paraphrases.
```

3) Marketing hypothesis generator

```
Role: growth strategist.
Context: goal — reduce CAC by 15% in a quarter.
Task: 20 hypotheses across channels (search, social, partnerships, content, email), prioritized by ICE and impact on CAC/LTV.
Format: [Hypothesis | Channel | Impact on CAC | Impact on LTV | ICE | First steps].
```

4) Funnel analysis: growth points

```
You are a funnel analyst.
Context: data: impressions=1 000 000, clicks=12 000, leads=360, meetings=170, deals=68.
Task: find bottlenecks, propose 10 hypotheses and 5 quick experiments.
Format: table [Stage | Current % | Benchmark | GAP | Hypotheses].
Model: YandexGPT/GPT‑6.x.
```

5) SQL query for BI (if you work with a DB)

```
Role: analyst assistant.
Context: tables events (ts, user_id, event), leads (lead_id, user_id, status).
Task: calculate CVR from "registration" → "lead" by week for 8 weeks.
Format: SQL + comments.
Constraints: do not execute, only show example query and checks.
```

—

## Prompt‑book for download: CSV and JSON (paste into your editor)

CSV version (prompt-book.csv):

```
id,task,goal,role,model,temp,format,constraints,acceptance,metrics
CNT01,Контент‑план,"Заполнить 4 недели контента","Контент‑стратег B2B","Claude Sonnet|GPT‑6.x",0.4,"Таблица [Неделя|Канал|Тема|Цель|CTA|KPI]","Без воды, РФ 2024–2026","3 кластера интентов, 30% репак","ER, CTR, лиды/нед"
SEO02,SEO‑статья,"Инфоинтент, охват","Редактор B2B","GPT‑6.x|Opus",0.3,"Markdown H2/H3 + таблица","Без ссылок, факты проверять","Структура, читабельность","Органический трафик, Time on page"
BRD03,Голос бренда,"Единый тон","Бренд‑редактор","YandexGPT|GigaChat",0.3,"v1/v2 + список замен","Без штампов","‑15% длина, +ясность","Редакт‑оценка"
SMM04,План SMM,"ER + клики","SMM‑редактор","YandexGPT",0.5,"Таблица 7 постов","Без мемов","3 хука/нед","ER, CTR"
ADS05,Объявления поиск,"+10% CTR","Редактор контекста","YandexGPT",0.4,"Таблица вариантов","Естественные ключи","3 механики","CTR, CVR"
ADS06,Креативы таргет,"10 концепций","Креативщик","GPT‑6.x",0.7,"Таблица идей","Без пустых обещаний","4/4/2 по риску","CTR, CPL"
CRM07,Welcome‑цепочка,"Демо","Email‑редактор","Claude|GPT‑6.x",0.4,"5 писем: тема/тело/CTA","Без спама","2 темы A/B","Open, CTR, CR"
CRM08,Реактивация,"Возврат","CRM‑маркетолог","YandexGPT",0.4,"3 письма + частота","Без давления","Опт‑аудитория","Open, CTR"
ANL09,JTBD‑карта,"Гипотезы боли","Исследователь","Claude|GPT‑6.x",0.2,"Таблица JTBD","Без домыслов","5 инсайтов","Кол‑во идей"
ANL10,Воронка,"Поиск узких мест","Аналитик","GPT‑6.x|YandexGPT",0.2,"Таблица этапов","Факты только","Точность расчётов","CVR, CPA"
```

JSON version (prompt-book.json):

```
[
  {
    "id": "CNT01",
    "task": "Контент‑план",
    "goal": "Заполнить 4 недели контента",
    "role": "Контент‑стратег B2B",
    "model": ["Claude Sonnet", "GPT‑6.x"],
    "temp": 0.4,
    "format": "Таблица [Неделя|Канал|Тема|Цель|CTA|KPI]",
    "constraints": ["Без воды", "РФ 2024–2026"],
    "acceptance": ["3 кластера интентов", "30% репак"],
    "metrics": ["ER", "CTR", "Лиды/нед"]
  },
  {
    "id": "ADS05",
    "task": "Объявления поиск",
    "goal": "+10% CTR",
    "role": "Редактор контекста",
    "model": ["YandexGPT"],
    "temp": 0.4,
    "format": "Таблица вариантов",
    "constraints": ["Естественные ключи"],
    "acceptance": ["3 механики"],
    "metrics": ["CTR", "CVR"]
  },
  {
    "id": "CRM07",
    "task": "Welcome‑цепочка",
    "goal": "Демо",
    "role": "Email‑редактор",
    "model": ["Claude Sonnet", "GPT‑6.x"],
    "temp": 0.4,
    "format": "5 писем: тема/тело/CTA",
    "constraints": ["Без спама"],
    "acceptance": ["2 темы A/B"],
    "metrics": ["Open Rate", "CTR", "CR к демо"]
  }
]
```

## How to test prompts: A/B, metrics and acceptance criteria

The test goal is to compare the lazy vs working prompt by business metric, not by subjective "like/dislike".

Typical metrics:

- Content/blog — time on page, depth, share of organic traffic, citation in neural search.
- SMM — ER (likes+comments / reach), CTR from post, referrals.
- Ads — CTR, CVR (click→lead conversion), CPA/CPL.
- Email — Open Rate, CTR, CR to the goal (demo/lead), unsubscribes.
- Sales funnel — stage conversion rates, cycle time, CAC, LTV.

Mini A/B checklist:

- Hypothesis with threshold: "New prompt reduces CPA by 15% at same lead volume." 
- Unit of randomization: ad / creative / email / landing block.
- 95% confidence level, fixed horizon or sequential testing.
- Equal budgets and targeting except for the variable.
- Duration — at least one funnel cycle (often 7–28 days in B2B).
- Stopping rule and decision criterion defined before start.

Example thresholds (table):

| Task                     | Main KPI    | Success threshold            | Minimum volume       |
|-------------------------|-------------|------------------------------|----------------------|
| Search ads              | CTR         | +10% vs benchmark            | ≥ 3 000 clicks       |
| Targeted ads            | CPA         | −15% at same lead volume     | ≥ 200 leads          |
| Welcome email #1        | Open Rate   | +5 p.p.                      | ≥ 10 000 recipients  |
| Telegram post           | CTR         | +12%                         | ≥ 50 000 impressions |
| Longread on blog        | Time on page| +20%                         | ≥ 1 500 sessions     |

Funnel before prompt optimization:

![Типовая воронка до оптимизации промптов: теряем лиды на каждом шаге](./images/prompty-dlya-marketologa-shablony-testy-bezopasnost-ab-funnel.png)

Common testing mistakes and how to avoid them:

- Peeking and premature stopping. Fix: lock horizon and confidence level.
- Mixing segments. Fix: keep audience and timing identical.
- Evaluating by a proxy metric. Fix: look at CVR/CPA, not only CTR.
- Poor formalization of the "working" prompt. Fix: save exact prompt text, system role and model params.

It helps to store before/after in one place. Average observed effect when switching to formalized prompts:

![Средний эффект по 12 A/B‑тестам в B2B и IT, 4 недели](./images/prompty-dlya-marketologa-shablony-testy-bezopasnost-effect.png)

## Model selection and settings for the Russian market

Short: choose the model by task, language and compliance needs.

![Оценка пригодности моделей под русскоязычный маркетинг (0–100, по прикладной практике 2026)](./images/prompty-dlya-marketologa-shablony-testy-bezopasnost-models.png)

Comparison (as of 2026):

| Task/criterion            | YandexGPT                 | GigaChat                  | GPT‑6.x                 | Claude Sonnet/Opus        |
|--------------------------|---------------------------|---------------------------|-------------------------|---------------------------|
| Russian copywriting      | Strong, natural tone      | Good, sometimes terse     | Excellent, but pricier  | Excellent, softer style   |
| Structural analytics     | Good                      | Good                      | Very strong             | Very strong               |
| Compliance/localization  | Easier with PD in RF      | Easier with PD in RF      | Transmission needs review| Transmission needs review |
| Cost/speed               | Competitive in RF         | Competitive in RF         | Depends on pricing      | Depends on pricing        |
| Where to apply           | SMM, copy, local tasks    | Email, copy, dialogues    | Research, strategy      | Research, longform copy   |

Default settings:

- Creative: temperature 0.6–0.8, top_p 0.8–0.9.
- Analytics/tables/SQL: temperature 0.1–0.3, strict formatting.
- Editing/styles: temperature 0.2–0.4, forbid "making up" facts.

When a local model or private perimeter is mandatory:

- If you have PD, commercial secrets, contractual limits, government contracts, or internal security policies.
- Solution: YandexGPT/GigaChat in a private contour or on‑prem LLM; or a pre‑saver with entity masking.

## Security and compliance: what not to send to public models

By law and corporate policy you must not send the following to public LLMs:

- PD: full names, phone numbers, emails, addresses, job titles, identifiers.
- Commercial secrets: contract amounts, margins, unannounced features/prices.
- Sensitive data: medical, biometric, political views, etc.

Safe prompt pattern (example):

```
You are a funnel analyst.
Context: [SECTOR]=distribution; [SIZE]=100–200 employees; Audience — ROP/CEO.
Task: generate 10 hypotheses to grow CVR at the "lead → meeting" stage.
Format: table [Hypothesis | Why | How to test | Risk | KPI].
Constraints: do not use PD; numbers replaced with tags [$CPA], [$CR].
Data (obfuscated): CTR=[$CTR], CVR=[$CR], CPA=[$CPA].
```

How to de‑identify:

- Mask: Ivan Petrov → [CLIENT_A]; 8 530 rub. → [$AMOUNT_A]; Moscow → [CITY_A].
- Aggregate: send summarized tables instead of raw logs.
- Separate secrets: do not send strategic documents to the cloud at all.

Security checklist before running a prompt:

- No PD or commercial secrets in text/data.
- Clear context and goal, fixed model and params.
- Output format regulated, data obfuscated.
- Quality criteria and KPI of the test defined.

## Common mistakes and how to avoid them

- No role and context. Fix: start with "You are…", define audience and JTBD.
- Vague task. Fix: one action per prompt; split complex tasks.
- Empty format. Fix: require a table/structure, specify length and fields.
- Missing constraints. Fix: ban fabrication, set style/tone, "no links".
- Testing by proxy metrics. Fix: lock KPI, focus on CVR/CPA, CAC/LTV.
- Sending PD to the cloud. Fix: local models, pre‑saver, obfuscation.

## Implementation checklist for the leader (CMO): stages, resources, ROI

Stages:

1) Diagnosis. Where AI is useful: content, SMM, performance, CRM, analytics.
2) Data policy. What can/cannot be sent. Local/private models for PD.
3) Prompt library. Create a single repository (like the prompt‑book above) with versioning.
4) Pilots. 2–3 directions, fixed KPIs (CTR, CPA, delivery speed).
5) Team training. Prompt formula, quality criteria, before/after.
6) A/B procedures. Unified counting rules, reporting, test calendar.
7) Scaling. Approve working templates as standards.

Resources:

- Process owner (marketing operations), quality editor, data analyst.
- 2–4 hours/week to maintain the library and run retros.

ROI (example):

- Content production saving: −30% time × team rate.
- Ad efficiency growth: −15–25% CPA at same volume.
- Model costs: track budget per 1000 tokens and share of local solutions.

## Quick before/after summary by category

Content:

- Before: "Write an article about CRM." 
- After: "Role: B2B editor, informational intent, H2/H3 structure, criteria table, 1800–2200 words, DoD."

SMM:

- Before: "Create a Telegram post." 
- After: "SMM editor, PAS structure, length, 3 hooks, CTA, ER/CTR metric."

Ads:

- Before: "Need texts." 
- After: "Contextual editor, 3 mechanics, BOFU, table, KPI CTR/CVR."

Email:

- Before: "Write an email." 
- After: "5‑email welcome sequence, A/B subjects, preheaders, CTA, Open/CTR/CR."

Analytics:

- Before: "Review competitors." 
- After: "Table of positions, gaps, ICE ideas, links to observed elements."

## Frequently asked questions (FAQ)

— duplicated in the structured FAQ block at the page bottom; here is a short version.

1) What to include in a prompt? Role, context, task, format, constraints, data, criteria and KPI. 
2) How to test? A/B on a business metric, 95% confidence, fixed horizon, stopping rule. 
3) Which models for Russian? Copywriting — YandexGPT/GigaChat; analytics — GPT‑6.x/Claude; PD — local models. 
4) PD to cloud? Not allowed; mask or use private perimeter. 
5) Channel adaptation? Keep context, change role/format (Telegram/landing/email).

—

Short reminder: prompts are a manageable process. Fix the formula, test on KPI, follow compliance — and AI will become your "fast second in command", saving time and lowering CPA without losing quality.

## FAQ

### What should I always include in a working prompt for a marketer?

Minimum: model role, context (market, audience, JTBD), a clear task, response format, constraints/quality criteria, source data and a success metric. Example: "You are a B2B editor. Context: service X, audience — CIOs. Task: draft a longread for informational SEO intent. Format: outline + 1500–1800 word text, CTA. Constraints: no fabricated facts, do not insert links. Criteria: readability, uniqueness, tone fit, 3 USPs."

### How do you test and measure prompt effectiveness?

Use A/B testing at the outcome level: for posts — CTR/ER, for lead gen — CR and CPA, for emails — Open Rate/CTR, for ads — CTR→CVR→CPA. Keep a 95% confidence level, document the hypothesis in advance, avoid peeking. Minimal duration — a full funnel cycle. Define the decision rule before start (for example, −15% CPA with unchanged lead volume).

### Which models are better for the Russian market and when is a local solution necessary?

As of 2026: for copywriting and dialogues in Russian — YandexGPT and GigaChat are strong and convenient for compliance. For complex structure/analytics — GPT‑6.x and Claude Sonnet/Opus. If you have personal data/commercial secrets — use local models or a private perimeter, or obfuscate data via a pre‑saver.

### Can I send client/personal data to public LLMs?

No. Personal data and confidential commercial information must not be uploaded to public cloud LLMs per legislation and internal company policy. Use a local LLM or a pre‑saver with masking: replace names/phones/amounts with tags like [CLIENT], [PHONE], [$AMOUNT].

### How to adapt one prompt for different channels (Telegram, landing page, email)?

Keep the same context and objective, change the role and format: "You are a Telegram SMM editor…" → short post + 3 hooks; "You are a landing copywriter…" → H1+sub+benefit blocks; "You are an email editor…" → subject+preheader+body+CTA. Specify length, tone and acceptance criteria (for example, no emojis for B2B).