The system that tells you exactly,
what to do next
I am Grigoriy Dobryakov — an engineer who knew nothing about marketing,
so I built a next-move AI engine that tells me exactly what to do next.
The Problem:
Marketing runs on opinion.
Engineering runs on data.
Most personal marketing is vibes: post something, wait, hope. No feedback loop. No signal processing. No prioritization logic.
I'm an engineer. I don't know how to work like that.
So I didn't.
without a system
- Post and pray
- Gut-feel prioritization
- No feedback loop
- Unknown signal-to-noise
- Can't explain decisions
with this system
- Classified signals daily
- Typed, executable actions
- State tracked in git
- Hypotheses with confidence
- Every decision explainable
The Solution:
A proper ETL pipeline. For market intelligence.
Five layers. Every layer answers one question and hands off to the next.
Layer 1
Inputs
Raw market data, collected daily
- · LinkedIn posts from targets
- · Facebook conversations
- · Job vacancies parsed
- · Inbound messages
Layer 2
Signals
Each input classified by type and strength (model on a fixed vocabulary)
- hiring-surge
- leadership-change
- public-complaint
- budget-signal
- tool-search
strength: strong / moderate / weak
Layer 3
Intelligence
Deep per-lead analysis — model-executed skills, structured outputs
- · Who is this person, real agenda
- · Pain behind the public signal
- · Claims fact-checked
- · Fit against products & cases
Layer 4
Hypotheses
Synthesized market model — model over versioned state
- · Macro — structural market shifts
- · Meso — company-level patterns
- · Micro — individual lead hypothesis
Each has a confidence score.
Low confidence → no outreach.
Layer 5
Actions
Concrete, typed, immediately executable
Action Items
who · channel · exact message · KPI · stop rule
Account Intelligence, not magic: drafts from the model; human sends or edits.
Positioning Items
which narrative to fix · what content to write
Dependency graph included.
AI-9 unlocks AI-10. No guessing.
What an action item actually looks like:
AI-1 — outreach action
- Who: Felix Mornebrik, Co-founder / CTO, Zalvernix
- Channel: LinkedIn DM or comment
- Angle: AI agent blast-radius without allowlists
- CTA: One closed question on tool contracts vs guardrails
- KPI: Any reply in 7 days
- Stop: Silence → log in people/, no retry
PI-5 — positioning action
- Type: Content positioning
- Hypothesis: Agent-readiness framing emerging
- Action: Write post on "agent-readable processes" as EM KPI
- Why now: C-level using this language publicly
- Metric: Post engagement + DM replies
Under the hood
Not a deck-and-framework person.
Here's the actual system.
Four system traces. Real file paths, real SKILL.md contracts, real decision rules. Follow any signal from raw input to typed output — every step is a versioned artefact in git.
trace 1
Vacancy → outreach brief
5 layers · 2 skills · 1 decision
read trace →
trace 2
PI action → published post
6 skills · evidence-first
read trace →
trace 3
4 weak signals → hypothesis
18-day arc · temporal synthesis
read trace →
trace 4
How skills hand off
SKILL.md contracts · file isolation
read trace →
This is a live system. Not a concept.
Current state of the market-state database, versioned in git.
21
buy signals
tracked
15
companies
in watch
14
hypotheses
synthesized
23
outreach
actions
11
positioning
items
git
full history,
every decision
AI didn't replace the thinking.
It replaced the legwork.
The system runs on a set of custom Claude Code skills. Each one is a specialist that handles one job and hands off to the next. Each has a SKILL.md — a spec with inputs, outputs, constraints, and quality criteria. Every invocation is a language-model run; the spec is what keeps the behavior bounded enough to trust.
| Skill | What it does | Layer |
|---|---|---|
| vacancy-intel | Read a job posting as a market signal — who's hiring and why, pain behind the requirements | Signals |
| lead-intel | Deep profile of an inbound lead — pain, agenda, fact-check of claims, fit against products & cases | Intelligence |
| market-analyst | Synthesize changes in market-state/ into micro / meso / macro hypotheses with confidence scores | Hypotheses |
| action-planner | Convert hypotheses into typed, executable AI (outreach) and PI (positioning) action items | Actions |
| action-update | Log status updates back into the right action file — finds relevant file, classifies input, updates state | Actions |
| evidence-finder | Pull proof points from cases and products before writing any content — always runs before publishing | Content |
| fb-post-writer | Write expert content posts with correct voice, structure, and audience calibration | Content |
| humanizer | Strip AI patterns before anything goes public — runs last, before every publish | Content |
| chronographer | Archive every published post automatically — no manual logging required | Archive |
Key principle:
The human decides strategy. The system executes and logs. Every output is versioned in git. Every decision is explainable.
The full surface:
Everything the system actually does.
Ten capability groups. Each card is a real, running function — backed by a versioned artefact in the repo, a skill with a frozen SKILL.md, or both. This is the breadth most people don't see from the headline pipeline.
See the market
Inbound lead breakdown
Any incoming text — email, chat, transcript — becomes a structured intel report with author profile, motives, and fact-check.
Vacancy as market signal
A job posting read for what sits behind it: pain, runway, org shift. Two lenses — buyer of consulting, employer of you.
Deep human profiling
Career arcs, downshift patterns, narrative drift, irrational triggers — not "what they do," but "what moves them now."
Compounding company profiles
One file per company. Every new signal lands in the same place. Financial, operational, narrative — all in one portrait.
Competitor & tech tracking
Direct rivals and the technologies shaping demand — each with its own card, updated on every meaningful signal.
Proof chain on every claim
"Someone said" vs "here are the source links." Every intel report carries links back to posts, filings, and public moves.
Synthesize
Insights — what you can't see by eye
The main thing the system makes for the owner. Not reports, not digests — conclusions a human couldn't assemble: too many weak signals, too short a window, too many sources.
Multi-signal convergence
Three independent layers — financial, operational, behavioral — pointing the same way. That's when confidence goes up and a window opens.
Pre-press-release prediction
Faint hiring + post + filing signals over 3–5 days yield a hypothesis before the announcement hits the news.
Reusable market patterns
One-off observations crystallize into patterns the system applies to new situations. PoC→production, compliance-as-precursor, frontier-lab-as-barometer.
Temporal analysis via git
State at any past date is recoverable. Each run diffs against the last tag — only the new gets processed. Trends fall out of the log.
Three altitudes
Macro shifts, meso company patterns, micro lead-level hypotheses — each dated, each with confidence, each checked on the next pass.
Decide & prioritize
Hypotheses → typed actions
Every hypothesis becomes concrete next steps. Two tracks: AI (outreach) and PI (positioning fixes).
Action windows with expiry
Not only "what to do" but "for how long it's still worth doing." 1–2 weeks for viral topics, 6–12 months for structural pain.
Next Move Engine
When signals converge, the system writes the brief: audience, angle, market signal, supporting data, target maturity. The market picks the topic.
Live feedback into actions
Reply from a decision-maker, status note, edit — free text routes itself into the right card without overwriting it.
Two-track pipeline
A signal may say "write this person" (AI) or "fix the segment narrative" (PI). Both tracks move in parallel.
Dependency graph
Actions know which other actions they unlock. No flat list, no guessing what's next.
Act precisely
Ready-to-send outreach copy
Not "reach out to X" but the actual message body — channel, angle, pain, proof point, CTA, all on the card.
Cartesian assembly
Role × segment × product × case × proof. Not a template with name substitution — a message built for this person's situation.
Maturity-driven routing
A CTO stuck on pilots and a CTO with agents in production speak different languages. The system picks the right one before writing.
Async-first CTA
Default cold-touch CTA is a comparative question answerable in one line — not "let's hop on a call."
S1 cold-touch playbook
Noticed a signal → careful read of what it means → light question. No CV, no metrics, no positioning dump in the first message.
Resume adapted per signal
A CV is an entry point into outreach, not a static doc. Built per target in MD and PDF.
Multilingual cold outreach
Language of first touch picked from the target's profile, not the sender's habit.
Produce content
LinkedIn writer
Posts that position for AI-driven engineering leadership. Algorithm-aware, maturity-aware, voice-locked.
Facebook writer (RU voice)
Provocative, expert, weighted. Four voice modes, indirect-sale mechanics, FB algorithm rules baked in.
Blog & Telegram outputs
Longer-form and channel-specific variants from the same source material.
Video via NotebookLM
Expert texts, cases, insights — converted to audio and published on YouTube. No separate scripting.
Reproducible content pipeline
brief → evidence → writer → expert-voice → humanizer → archive. Skills with explicit contracts, not vibes.
AI-pattern removal
Strips slop, rhythmic flatness, cliché structures — runs last, before every publish, even if the text already sounds fine.
Expert-depth calibration
Cures curse-of-knowledge spots where an expert talks about complex things as if obvious. Minimal edits, voice preserved.
Evidence base from repo
Every publish is grounded in real cases, numbers, and quotes from the repository — no invented metrics.
Silent post archive
Every produced post is saved automatically, named by platform and topic, dated, filed — no manual logging.
Validate positioning
Customer interview simulation
Test a pain hypothesis in dialogue before going public. Persona pulled from decision-makers, segment, maturity model.
Business idea red team
Not just "rate this idea." Mandatory critique: market, competitors, monetization, channels, risks — and fit with current positioning.
Positioning fit check
Every intel report and hypothesis is auto-cross-referenced with products, segments, cases — gaps surface immediately.
Accumulate expertise
Personal book library via MCP
Authoritative sources at hand. Search, read, index — through an Obsidian MCP server, not a generic search engine.
Disciplined RAG protocol
Routing by task domain, hybrid mode, relevance threshold, mandatory full-source reads — not snippets and vibes.
Methodologies in knowledge-base
Not raw quotes — structured methods organized by task: channels, semantic consumer, EM implementation.
Silent knowledge extraction
Every substantial conversation leaves a trace — new terms, patterns, concepts, names. Like autosave for thinking.
Dialogue archive + search
Decisions, corrections, agreements logged. Searchable: "how did my thinking on X evolve?" returns a chronology with sources.
Public face
Bilingual site from source
Site is a consequence of operations, not a separate asset. Pipeline: products/*.md → content.md → index.html. One deploy command.
Cross-cutting content edits
When positioning shifts, the AI rewrites not just new copy — page hierarchy, internal links, product descriptions, even old pages get re-voiced.
Site ↔ analytics closed loop
GA/YM metrics feed back as a signal. Positioning change → site diff → behavior change → next iteration.
Measure & learn
Real post-statistics analysis
Not likes-vs-no-likes — correlations between content moves and response across the archive. Outputs actionable rules for the next brief.
Feedback into next brief
A post with 105K views and 203 comments influences what gets generated next — statistics flow into the Next Move Engine as a factor.
Evolution reconstruction
Track how a view changed over time via git and dialogue logs. Question in, dated trail out.
Under the hood
ETL as marketing metaphor
Extract inputs, transform against state, load into channels. Feedback returns to input. Marketing as an engineered service, not a creative practice.
inputs / state / outputs at every step
Each agent sees only what it needs. Raw inputs, structured living state, concrete outputs. Microservice analogy holds.
Git as a temporal layer
Time is a built-in dimension, not a feature. Tags mark agent runs, diffs feed the next pass, history is the trend source.
Skills with SDLC
Not prompts — versioned modules. Reference texts for eval runs, explicit contracts between skills, versioned outputs.
Scheduled agents
Part of the work happens with no human in the loop. Periodic agents run on cron and refresh state.
MCP integrations
The system talks to external sources over a standard protocol: Obsidian books, vacancies, scheduled tasks, browser, session management.
The headline pipeline is just the spine.
Around it sits everything above — intel, synthesis, validation, content, measurement, infrastructure. That is what makes the next move trustworthy enough to execute without second-guessing.
The lesson
"If you know how to design systems,
any domain is just the schema."
The domain was unfamiliar.
I hadn't built a marketing system before. I have built data pipelines, event-driven architectures, and feedback loops.
The problem was the same.
Raw events need classification. Patterns need to surface from noise. Decisions need to be based on state, not intuition.
It took two weeks to go live.
From zero to operational. It now runs daily. Every output has a success criterion and a stop condition.
Now I just run the next move.
Every day starts with one ranked action from the system. I execute, measure, and feed results back into the loop.
Grigoriy Dobryakov
Engineering Manager. 25+ years in the industry. Now building with AI.
Engineering Manager with a background in delivery, architecture, and team building across startups and enterprises. Led engineering at Askona (15M customers, 20+ distributed teams), UMI (50% revenue growth), Sprinthost, Personaclick, and Korus Consulting.
This project started as a side experiment: can I apply engineering discipline to my own positioning? It became a working system I run every day.
Want the same for your business? I can build your next-move AI system.
Active signals by type
Latest hypothesis
Macro · moderate confidence
C-level IT leaders in transition are producing AI narratives on LinkedIn — not just content, but a signal of positioning for next role or consulting market.
→ Generated 2 outreach + 3 positioning actions
Want to talk engineering, AI, or systems thinking?
I'm available for conversations about engineering leadership, AI-augmented workflows, and delivery systems. Not pitching. Just talking.