Method · AI orchestration · Pipelineniche-independent

Not a case — an operating system

The same pipeline replaces a team on research, writing and checking. I hold the truth and the decisions, while the orchestration researches, designs, produces and adversarially checks. The niche sets the content, but doesn’t change the pipeline.

145
commits
7
repositories
10
calendar days
200+
AI agents

доказаноevery number is checkable in code and gitкак проверяем →

Role boundary

Who owns what

This isn’t “I hired a team” — it’s “I designed a pipeline that replaces a team.” The boundary is hard and held on purpose: the niche sets the content, but doesn’t change the pipeline. It all starts with who is responsible for what.

The human owns
Truth · face · decisions
  • Truth — what’s true about the product and the niche.
  • Face and brand — who’s speaking.
  • Decisions — what ships, what’s out of scope.
  • The final yes to ship — the last word stays with the human.
The orchestration owns
Production and checking
  • Research — crawl, parse, synthesise.
  • Design — architecture, schemas, contracts.
  • Production — code, copy, layout, assembly.
  • Adversarial review — an attack on its own result.

The boundary is the portable asset. It removes the main bottleneck of working solo — the wait for sign-offs: the human doesn’t produce or proofread line by line, and the orchestration doesn’t decide what’s true about the business.

Six phases

From idea to autopilot

The pipeline runs through six phases. Each has the abstract work plus one checkable fact from a real run: proof the phase happened on the project, and didn’t just stay on paper.

  1. 0

    Standards — the frame before the plan

    A code of operating standards: who’s responsible for what, what’s out of scope, what’s flexible. “Invented constraints” are banned. The orchestration checks against the code before every plan.

    6 standards every plan returns to
  2. 1

    Research — field recon of the niche

    Competitors, funnels, price ladders, a catalogue of manipulative tactics — for inversion. Facts with sources, not averaged industry numbers. A number with no saved source is flagged “not confirmed.”

    9 competitors · 785 links · 42 verified prices
  3. 2

    Strategy — the plan and the ethics at once

    Positioning, channel design, the production rhythm, transparent prices as an advantage over obfuscation. The ethical frame is fixed here, not glued on later.

    7 key documents + 1 brand voice across every surface
  4. 3

    Assembly — the whole funnel at once

    Every layer of one funnel is built in parallel: the core product, the content product, the pages, the entry point. The layers are linked by contracts — not “product first, then marketing.”

    6 products and subsystems in one funnel
  5. 4

    Quality gate — to production or back for rework

    Before shipping, every block passes a gate with one of two outcomes. Security isn’t a final checkbox but a veto loop in every cycle. Code checks: types → style → unit tests → build, stopping on the first error.

    book audit: 62 agents → 77 findings
  6. 5

    Autopilot — the system runs itself

    Scheduled mailings, lifecycle notifications, a timezone-aware scheduler, a progress log. Content goes out in batches through a single brand filter. Edit the source — and the re-assembly is deterministic.

    Telegram bot: a state machine over 7 states · 16 touchpoints
Adversarial cycle

Every block is attacked, not reviewed

This is the core of the quality gate — and the main thing that sets the pipeline apart from “generate and forget.” Swarms of agents research, write and assemble, while I hold the truth and the decisions. The key: the check isn’t sequential but adversarial — every block is attacked, not “glanced at.”

cycle · repeats on every block
1

Draft

An agent writes the block.

Three veto lenses∥ in parallel— each in an independent context
L1 · fact accuracy“attack the assumption,” not “check it”
L2 · safetyroute to a doctor, not a protocol
L3 · voice and structurebrand tone, block logic

Revision decision

↺ back into the cycleto production → integration
after integration · a separate run
4th lens · “break this”

A fresh, independent agent hits the already-assembled block. Its only job is to break it. This isn’t a step of the cycle but a separate final break-in after integration. The author’s and the reviewer’s errors are independent.

The scale of one adversarial run — a book audit:

62agents77findings53confirmed1false

Through the content cycles 200+ agents have run; each with a real wf_*-id, outputs structured — a machine-readable base of facts, not prose.

The orchestration engine

Four supporting mechanisms

Adversarial review is above, as the centre of the method. Under it run four more mechanisms that make the pipeline a pipeline. All of them are procedural — they port into any niche.

Agent fan-out

A task is split into dozens of agents with explicit contracts: what it reads, what it must return, in what form. Scale grows with the number of tasks, not the people hired.

Structured output

Agents return machine-readable structures (findings, contracts), not prose. The outputs are checked programmatically, so the “to production or back for rework” call rests on data, not on an impression.

Docs-loop memory (5 zones)

Source of truth · active work · audits (frozen snapshots) · reference · archive. A “last checked” mark, a plan log. A new agent enters the task with the same context.

Resume and re-assembly

Against the rate limit: not a retry loop, but recovery and re-assembly from the log. Already-computed drafts and lenses are reused. The rate limit becomes a pause.

Portable techniques

What ports into any niche

The eight strongest techniques in the pipeline. Each comes with what it is concretely and where it ports. The niche changes the content of a technique, but not its form.

Meta-repository

An umbrella repo with nested submodules, each pinned to a commit. One umbrella commit is a version snapshot of the whole ecosystem; a rollback or an audit takes one command.

Ports to: any system of several linked products.

Docs-loop memory

Five zones: source of truth, active work, frozen audits, reference, archive. Each audit has a checked-on date, each task a log, and a new agent enters with full context.

Ports to: any work longer than one session — the foundation of autopilot.

Agent fan-out

A task is split into dozens of parallel agents, each with an explicit contract: what it reads, what it must return and in what form. Scale grows with the number of tasks, not the people hired.

Ports to: any task divisible into independent subtasks.

Adversarial review

Several independent lenses tasked with “find what’s broken,” and a final fresh agent to “break this.” The author’s and the reviewer’s errors are independent, so they’re caught before production.

Ports to: any pipeline shipping content or code to production.

Structured output

Agents return machine-readable structures — findings, contracts — instead of prose. The outputs are checked programmatically, so the “to production or back for rework” call rests on data.

Ports to: any pipeline where outputs have to be assembled and compared.

One source, many editions

One source of truth deterministically unfolds into several formats with no copying: a book → PDF, EPUB and a cut; product settings → landings and offers.

Ports to: documentation, content products, multi-format releases.

Data protection at the core

Versioned consent to collection, an immutable action log and deletion on request both ways — the app and the bot. Built into the schema, not bolted on before launch.

Ports to: any product with personal data; 152-FZ as a head start on GDPR.

Static at the core

Every route is statically pre-rendered; client JS is added as targeted islands only where interactivity is needed, not by hydrating the whole page.

Ports to: marketing pages where load speed and control over client JS matter.

The techniques port because the load-bearing structure of the pipeline doesn’t depend on the content. It rests on three pillars:

  • Separation of roles — the human holds the truth and the decisions, the wait for sign-offs disappears.
  • Gates instead of trust — adversarial lenses and an audit on content, checks on code, a binary “to production / back for rework.”
  • Memory that outlives the session — docs-loop memory and the umbrella repo hold state between runs; without it, autopilot is impossible.
Method boundaries

I claim only the provable

Honesty about its own boundaries is part of the method. It carries a system to production and does so reproducibly; but there’s also what the method by itself doesn’t prove. I separate the two outright.

What the method proves
  • Carries a system from idea to production — measurable in git: 145 commits in 10 days, each run reproducible by its wf_*-id.
  • Errors are caught before production: the author’s and the reviewer’s are independent, plus a final “break this” agent.
  • Security and data protection are part of the schema from day one, not a patch before launch.
Estimate and limits
  • оценкаPorting across niches — so far one run; the “80 / 20” is an estimate, not statistics yet.
  • не заявляюMarket results — the method carries to production, but a product sells, not the method; the first cohort has only just started, so I don’t claim revenue or customers.
  • не заявляюThe 1→N hardening — load, end-to-end tests, real-user shakedown: that’s already separate work on the cohort.

Designed → built → verified → shipped to production.

Frequent questions

What people usually ask

Next

Check the method against the facts

The method is proven on a real run, ports into another niche and is all reduced to checkable artifacts. Three doors — each answers a different question.

Let’s design your pipeline

The same method ports into another niche. Write to me — we’ll work out what here gives you the lever.