Dmytro Zubchenko Book a call

Dmytro Zubchenko AI and automation

Marketing AI systems with Claude Code and MCP

Dmytro Zubchenko has built six marketing AI systems with Claude Code and MCP: a content engine, daily paid-media reviews, a weekly marketing deck with daily plan-vs-actual, an ads-to-CRM attribution layer, an autonomous publishing loop, and custom integrations (about 10 MCP servers and wrappers, 19 custom agent skills).

Marketing AI systems that automate routine work and improve the quality of decisions. Not prompts in a chat window, but scheduled agents, custom MCP servers and gated pipelines that pull real data, do the work and show their evidence. I designed and built every one of them.

The engine room

The six systems: what each one does and where it runs

Each system below shows its pipeline, what goes in, what comes out and the guardrails around it. Every figure is one I documented from the runs themselves.

Systems
6systems, documented below
MCP servers
~10MCP servers and wrappers authored
Agent skills
19custom agent skills
Decks
15weekly decks, June to October 2026
Fabricated claims
0fabricated claims in a 25-article release, per the pipeline's own quality gate

Content engine

PitchMonster · Sales Bureau

A topic or a live URL goes in. A gated, publish-ready article comes out. Two human approval gates and a 0-100 quality gate with hard stops sit in between.

  1. Approved content planTopic or live URL
  2. Keyword brief
  3. SERP and AI-citation teardown
  4. Human approvalBefore drafting
  5. Draft
  6. SEO, schema and polish
  7. Quality gate0-100, hard stops
  8. Human approvalBefore publishing
  9. Staging-first publish
  10. Render and sitemap verificationPublish-ready
Runs at
PitchMonster, Sales Bureau
Input
An approved content plan: a topic or a live URL.
Output
A gated, publish-ready article, checked by render and sitemap verification.
Guardrails
Human approval before drafting and before publishing. A 0-100 quality gate with hard stops for a fabricated claim, an invented link or a missing approval.
Stack
Claude Code skills
DataForSEO
Search Console
GA4
WordPress adapter
Webflow adapter
Payload adapter

Hard stops

  • StopFabricated claim
  • StopInvented link
  • StopMissing approval

Documented scale

25articles
141,000+words
318images
0fabricated claims, per the quality gate

One PitchMonster blog refresh release. At Sales Bureau, 29 of 30 planned articles are published.

Daily paid-media reviews

Sales Bureau

Every weekday the system pulls both ad platforms, joins spend to CRM rows and paid invoices, and reviews the account. It is allowed to change one thing on its own: high-confidence negative keywords, after a conflict check.

  1. Scheduled runWeekdays
  2. Pull Google Ads and Meta
  3. Join to CRM rows and paid invoices
  4. Classify
  5. Conflict check
  6. Auto-apply negative keywordsHigh-confidence only
  7. Re-read the account
  8. Written report and recommendation ledger
Runs at
Sales Bureau
Input
A scheduled run every weekday.
Output
A written report and a recommendation ledger.
Guardrails
The only change it makes on its own is high-confidence negative keywords, after a conflict check.

PrincipleCampaigns are judged on CRM revenue, not on platform-reported conversions.

RuleA failed pull is not a zero.

Weekly marketing deck and daily plan-vs-actual

Sales Bureau

One data file feeds a generated slide deck. Every number traces to one source, and revenue comes from paid invoices.

  1. Scheduled runDaily
  2. Pull Google Ads, Meta and CRM
  3. One data file
  4. Generated HTML slide deck
  5. Automated check
  6. Weekly reviewHuman
Runs at
Sales Bureau
Input
A scheduled run every day: Google Ads, Meta and CRM data into one data file.
Output
A generated HTML slide deck, reviewed by a person each week. 15 weekly decks from June to October 2026.
Guardrails
Every number traces to one source, revenue comes from paid invoices, and an automated check runs before the human weekly review.

Documented scale

15weekly decks, June to October 2026

Attribution layer and creative-level revenue reporting

Ads × CRM

Ad delivery is joined to CRM payments by click IDs and UTMs, so each creative is reported on the revenue it brought in.

  • Ad deliveryBy creative
  • CRM payments
  1. Join by click IDs and UTMs
  2. Tier creatives by sample size
  3. Creative-level revenue report
Runs at
Ads and CRM
Input
Ad delivery by creative, and CRM payments.
Output
A creative-level revenue report: each creative is reported on the revenue it brought in.
Guardrails
Creatives are tiered by sample size. Never let a number look more certain than it is.

RuleNever let a number look more certain than it is.

Autonomous publishing loop

PitchMonster

A scheduled headless run picks the next card from the task board, writes or refreshes the article with the content engine, and verifies the result in production before it pings the search engines.

  1. Scheduled headless run
  2. Pick the next cardTask board
  3. Write or refreshContent engine
  4. Machine gate and second-model review
  5. Cover image
  6. CMS import
  7. Production verification
  8. Search-engine index ping
Runs at
PitchMonster
Input
A scheduled headless run that picks the next card from the task board.
Output
An article written or refreshed with the content engine, verified in production, then a search-engine index ping.
Guardrails
A machine gate and a second-model review. One article per run, a lock file and a dry-run mode. Each run finishes and verifies one article before the next can start.
Safety rails
One article per run
Lock file
Dry-run mode

DesignA loop, not a cadence: each run finishes and verifies one article before the next can start.

Custom integrations

MCP · skills · routines

The systems above run on connectors I wrote myself: MCP servers and wrappers that give Claude Code direct access to the marketing stack, plus custom agent skills and recurring scheduled routines.

GOOGLE MKT GOOGLE ADS LINKEDIN ADS MAKE TAG MANAGER PROJECTS HR CLAUDE CODE 19 ORBITING DOTS = 19 AGENT SKILLS
Servers
~10MCP servers and wrappers
Tools
21tools in the Google marketing server
Skills
19custom agent skills
  • Google marketing server Search Console, GA4, Tag Manager
  • Google Ads
  • LinkedIn Ads
  • Make
  • Tag-manager server
  • Project-management connector
  • HR connector
  • Recurring scheduled routines
Covers
MCP servers, skills and routines
Gives Claude Code
Direct access to the marketing stack for Claude Code: about 10 MCP servers and wrappers (21 tools in the Google marketing server), 19 custom agent skills and recurring scheduled routines.
Role
These are the connectors the other five systems run on.

Operating principles, in every system

Human approval gatesPeople approve before drafting and before publishing.
Evidence before opinionThe data is pulled and read before anything is recommended.
Never fabricateA fabricated claim or an invented link is a hard stop.
Revenue from invoicesRevenue comes from paid invoices, not deal fields.
Model-tier cost controlSmall models pull data, larger models judge.
Dry-run, lock files, rollbackAutomated runs can be rehearsed, locked and rolled back.

Where they run

Where the systems run: PitchMonster, Sales Bureau and across accounts

Head of Product Marketing, part-time

PitchMonster

AI role-play platform for sales training.

Contact

Hiring a Head of Marketing or Head of Growth?

Open to Head of Marketing and Head of Growth roles, full-time or fractional. Tell me the stage, the team and the number that has to move, and I will tell you how I would approach it.

dmitrizubchenko@gmail.com · Krakow, Poland · Remote