Related: For the practical Claude team manual, see How to Claude — SEO Navigator Edition · This doc = Dennis Yu's task library methodology · That doc = Claude-specific team playbook Also: AI driven agency (strategic centaur agency vision & 7 pillars).
Methodology Guideline

Dennis Yu's "Look Over My Shoulder" AI Agents Marketing System

A framework for using AI agents to execute real marketing work — writing articles, publishing blog posts, building knowledge bases, running SEO audits — using the Task Library, Definitive Articles, QA Checklists, and the 4-Stage Content Factory.

1. The Core Principle

AI agents are not chatbots. They don't answer questions — they execute tasks. The difference is that every task is backed by documented methodology: how you do it, proof that you've done it, and a checklist to verify it was done correctly.

The golden rule: Never tell an agent "do the task." Always say "do the task according to [your documented methodology]." Without the invocation, the agent does whatever random thing it wants. With it, the agent follows your exact process.

2. The Task Library

The Task Library is your complete inventory of every repeatable task in your business. Think of it as the table of contents for how your company operates.

Examples of Tasks

Each task in the library points to a Definitive Article — the single-source-of-truth document that defines exactly how that task is done.

3. Definitive Articles

A Definitive Article is an SOP on steroids. It's the step-by-step walkthrough of a specific task with all supporting context baked in.

Anatomy of a Definitive Article

  1. Prerequisites — Access requirements (video storage, CMS, CRM, tools)
  2. Associated Training — Background knowledge the executor needs. "Raw ingredients" — without this context, you're generating garbage.
  3. Step-by-Step Process — The actual procedure with visual diagrams where helpful
  4. Examples — Multiple examples showing how the task is done correctly and incorrectly
  5. QA Checklist — A verification list to confirm the work meets standards
  6. Skill.md File — A markdown file formatted for AI agent consumption so the agent can follow the same process
Dennis makes all Definitive Articles public rather than locked in private skill files. The reasoning: constantly updating and redistributing private files is unsustainable. Public documentation lets anyone use it, improve it, and the AI can always find the latest version.

4. Meta Articles

After a task is completed, a Meta Article documents how it was done. This is the proof layer — evidence that the task was actually performed and what the results were.

A Meta Article Contains

Meta Articles close the feedback loop. The data from completed tasks feeds back into the knowledge base, updating how future tasks are executed.

5. The Invocation Pattern

When you give an agent a task, the structure matters. Dennis uses a consistent invocation format:

According to [your methodology/guidelines]
do [the specific task]
and QA your work using the checklist
then write a meta article explaining how you did it
and publish to [destination]

Real Example from the Video

"Write an article about my trip to Turkey visiting the 12 churches in Revelation. Look at my Google Photos for pictures tagged with Danny Leibrandt in Turkey. Write it according to the BlitzMetrics article guidelines and QA your work. Make sure it passes the check. Publish on dennisu.com. Tag friends from the journey. Reference related travels. Write a meta article according to the BlitzMetrics meta article guidelines. Cue up social media posts to promote it."

Notice: one voice prompt, ~2 minutes of speaking, triggers an agent to execute dozens of sub-steps autonomously.

6. The QA Layer

Never trust an agent that says "I'm done." Every task requires verification.

Self-QA

The same agent checks its own work against the QA checklist embedded in the Definitive Article. Built into the invocation: "and QA your work."

Cross-Model QA

Use a different model to police the work. Dennis's preference: Claude does the work, ChatGPT does the QA. Each model has strengths — use them accordingly.

The Postmortem Step

After QA, ask the agent: "How much of my guidelines did you use vs. outside sources?" A typical answer might be "80% your methodology, 20% Google's Quality Rater Guidelines." This transparency lets you decide whether to tighten constraints or allow the agent's judgment.

7. Building a Knowledge Base

Before agents can do meaningful work, they need raw material — your DNA as a business or individual.

Sources of Knowledge

Zoom calls Podcast episodes YouTube videos Email history Customer reviews Google Business Profile LinkedIn recommendations Speeches & talks Articles & blog posts Photos (geotagged) Social media posts Internal SOPs

The Topic Wheel

After collecting all positive mentions, the AI organizes them into a Topic Wheel — a map of what you or your company are known for. Each spoke of the wheel has associated content and people (relationships that drive trust).

The Topic Wheel feeds into the Knowledge Graph — the structure that earns knowledge panels and establishes topical authority in both traditional search and AI search engines.

Topical Authority = Relevant Platform + Matching Content

Content must be published where it's topically congruent. Digital marketing content goes on digital marketing sites. An ebook about remodeling goes on an ebook platform. A personal faith story goes on a personal site. Google measures congruence — a roofing article on a marketing site doesn't work.

8. The 4-Stage Content Factory

All marketing breaks down into four stages. If you're doing it right, you do Stage 1 and AI does Stages 2–4.

1
Produce
You create real experiences — field work, customer interactions, meals with friends, photos, reviews
2
Process
AI agents trim, transcribe, extract hooks, organize, and write from raw material
3
Post
Agents publish to your CMS, social platforms, and content distribution channels
4
Promote
Agents run ads, create social posts, syndicate, and amplify across channels
Your leverage is in Stage 1. Be on the roof. Be with the customer. Have the dinner. Take the trip. That signal — photos, reviews, conversations — is what the AI can't manufacture. Everything downstream is processing, and AI is better at processing than you are.

9. Operational Setup

Model Allocation

Claude (Opus) Deep work — writing code, sustained article writing, complex analysis, implementation
ChatGPT QA, document formatting, verification, policing Claude's output
Perplexity Overflow/overage work when Claude's usage limits are hit, research queries

Multi-Agent Browser Setup

Dennis runs ~10 agents simultaneously across browser tab groups. Each agent operates in its own set of tabs (2–4 tabs per agent) inside a Chromium-based browser. His personal work happens in a separate browser to avoid agents hijacking his active tabs.

Cost Structure

On the $200/month Claude Max plan with a 20x limit. When usage exceeds the plan cap, extra usage runs roughly $1/minute with 10 concurrent agents. Strategy: do heavy lifting within the plan limits, use Perplexity for overflow.

10. The Recursive Learning Loop

This is what makes the system compound over time rather than just repeat.

  1. Task Execution — Agent completes work following the Definitive Article
  2. QA — Work is verified against checklist (self + cross-model)
  3. Meta Article — Documents what happened, decisions made, data collected
  4. Data Loop — Real business metrics (calls, conversions, revenue) feed back
  5. Training Update — Agent suggests refinements to the methodology based on results
  6. Knowledge Base Evolves — Updated SOPs, new examples, better benchmarks

The agents learn from their own output. The system self-improves. This is the difference between a static SOP and a living, recursive methodology.

Quick-Start Checklist

Based on Dennis Yu's "Look Over My Shoulder: My AI Agents Do Marketing" — Marketing Mechanic Series
blitzmetrics.com · dennisu.com