Lesson 1 · 15 min
An AI agent is good at repetitive work: researching, drafting, checking, alerting, reporting. It is bad at deciding what is worth publishing. The inemaSEO method's machine keeps the two apart: the agents work in a line, and two gates stop everything until you approve.
Start the lesson
Goal: build the researcher → observer line with two human gates and an AGENTS.md that says what the agent never does on its own.
Tap each block.
The agents do the volume of work; the two decisions that cost the most stay with you.
Tap each step.
The method says to run the model through the subscription (Claude Code or Codex CLI) whenever possible. A paid API call only with explicit authorization for that purpose.
That way the cost is predictable, and no key becomes a shortcut for the agent to spend on its own.
The seven roles can be the same agent, given different requests. What matters is the order and the gates, not how many programs are running.
Start with a single role (the auditor, in lesson 2) and add the others once the previous one is reliable.
A gate is not "the agent tells you and keeps going." The agent stops, hands over what it did, and only continues after your yes.
At INEMA, the drafts for pillar pages 2 and 3 sat at gate 2 (about 3,150 and 2,450 words) and were only published on 9/24, after review.
Anything that leaves the site (push, notice to a search engine, email) or that spends money goes through a gate.
Claude Code and Codex read an instructions file at the root of the project before they work (Claude Code reads CLAUDE.md; Codex reads AGENTS.md). Write down there what the agent does, what it never does on its own, and how it works in batches.
Every fixed failure becomes a new rule in that file. Lesson 8 shows the real failures that became rules.
Copy the machine into the project README and the AGENTS.md into the root. Replace whatever is between the angle brackets.
Human picks the topic of the month ├─ Researcher → keyword list with score (autocomplete, Trends, Search Console, SERP) ├─ Opportunity → 5–10 GO with the URL and the real asset that provides the first-hand data │ GATE 1: human approves the list ├─ Briefing → intent, top 10 gaps, required first-hand data, internal links, link ladder ├─ Writer → draft that reads the repository/the transcript of the real asset ├─ QA → checklist (docs/CHECKLIST.md) + tests + link ladder │ GATE 2: human reviews (10–20 min per page) ├─ Publisher → commit + push, notice to search engines, post-deploy check └─ Observer → weekly export, what went up, what got stuck → back to GATE 1
# AGENTS.md — SEO for <yoursite.com> ## What you do - Research, briefing, draft, QA (docs/CHECKLIST.md) and weekly report. - Everything through the subscription (Claude Code or Codex). No API without authorization. ## What you never do on your own (stop and ask) - Call an API (Search Console, Bing, IndexNow, models) without written authorization for that purpose. - Commit + push a new page: only after GATE 2 (I review every page). - Delete a page, file or URL that has already been indexed. - State that something does not exist without checking (e.g., Search Console: look at the DNS TXT record first). ## How you work - Batch: dry-run first; marker to avoid duplicates; skip a repository with uncommitted changes. - Text from a source: only facts from the source. An item that fails the check is discarded, not fixed. - Another session may touch the same repo: commit only your own files; git pull --rebase --autostash before the push. - Secrets stay in .env (which is in .gitignore). Only the variable NAME goes here. - A fixed failure becomes one line in FALHAS.md: date | what broke | smallest fix | prompt or infra.
One role at a time, only after the previous one is reliable.
Use your own site or a test site.
Create the SEO project folder, copy the AGENTS.md (or CLAUDE.md) template and adapt it to your site. Open Claude Code or Codex in the folder and ask: "which actions can you not take without asking me?".
Done when: the file is at the root of the project and the agent answers citing at least the rules about API, push and delete.
Pick a topic of the month and write, for each role in the machine, what it delivers and where the result is kept (file, folder, spreadsheet).
Done when: you have the seven roles with what each one delivers, and the two gates marked with who approves and in which file the approval is recorded.
The explanation appears after your answer.
Question 1 of 4
Where are the two human gates of the machine?skill
Question 2 of 4
The agent found an API key in the project. Can it use it to notify the search engines?
Question 3 of 4
How much time does the method expect for gate 2?
Question 4 of 4
What is AGENTS.md (or CLAUDE.md) at the project root for?
Choose an option. The explanation appears after your attempt.