Strategy

I let a Claude agent run rpsg.co.id. Here's the setup.

How I handed my 6-month organic search plan to a Claude agent. The architecture, the autopilot crons, the 4-layer fact-check, what I still own.

3 Jun 2026 · 6 min read · by Ridho Putradi S'Gara

Most "AI runs my SEO" demos stop at "ChatGPT writes my blog post." This experiment goes further. I handed Claude my entire 6-month organic search plan for rpsg.co.id and asked it to manage execution end-to-end. Here is what it built, what surprised me, and what I still keep my hands on.

TL;DR

The starting point

This site is a brand-new domain. May 2026 first posts, near-zero authority, no referring domains worth counting. The 6-month plan calls for foundation in M1, topic clusters in M2, first links in M3, commercial intent and local in M4, AI citation push in M5, compound and systematise in M6. One quality post per week. The plan itself lives in a Google Doc.

My setup is unremarkable: an Astro site on Firebase Hosting, a Sanity studio I rarely log into, a Notion workspace already crowded with other projects, and a GitHub Action that publishes posts via the Anthropic API.

What I wanted to know: how much of the plan can an agent execute without me hand-holding every step?

Step 1, a command center in Notion

The agent's first move was to set up its own state. Six databases parented under my existing SEO hub:

Plus a single-row "agent lock" page used as an advisory lock against concurrent writes.

Forty-eight plan-tracker rows seeded. Ten priority content ideas in the queue. Twenty-five candidate queries proposed, of which I picked the final 20. That took about ten minutes of my time.

Step 2, subagents and slash commands

Claude did not get one giant prompt. It split itself into six subagents, each with a focused job:

Fourteen slash commands sit on top of those agents: status, weekly-prep, citation-audit, outreach-batch, monthly-report, and so on.

Three skills carry the voice rules: forbidden phrases, the case-study allowlist, the structural rules every post must follow. The same rules apply to blog posts, pitch emails, and anything else written in my name.

Step 3, technical SEO that actually crawls

This is the part most "AI SEO" tools skip. The agent connects to a real Screaming Frog crawler through an MCP integration. Every month on the 22nd, it crawls rpsg.co.id and reports back on:

When an orphan shows up, the agent files a refresh item in the content queue automatically.

Asking a chatbot to "analyze my sitemap" is not the same as a real crawl. The difference shows up when you have an actual broken link and no agent in the loop catches it.

Step 4, autopilot crons

Six scheduled tasks run unattended in Jakarta time:

The Wednesday publish itself runs as a GitHub Action. It pulls the next scheduled content queue item, hands the brief to Claude, runs four layers of validation, publishes to Sanity, builds, and deploys to Firebase Hosting. End to end without me in the loop.

Step 5, anti-hallucination guardrails

This is the layer I would not ship without. Four checks run before any post is published:

  1. Vague-citation phrase blocklist. A regex pass over every sentence catches authority-by-vibes patterns: appeals to unnamed studies, unnamed experts, unnamed industry surveys, and generic-population percentages. Any of those triggers a fail unless the same sentence has an inline link to a primary source.
  2. Unknown-metric detection. Any numeric claim about an entity that is not on the case-study allowlist, not framed as first-person experience, and not inline-linked.
  3. External link reachability. Every external link is fetched with a 10-second timeout. 4xx or 5xx fails the post.
  4. Adversarial fact-check. A second Claude call with web search categorises every claim as supported, first-person-acceptable, allowlist-match, common-knowledge, or unsupported. Any unsupported claim kills the post.

If any layer fails, the script exits non-zero. Sanity is not written. The week's slot stays empty. A missed Wednesday beats a fabricated post.

This guardrail is stricter than most human editors I have worked with. That is intentional. The whole point of rpsg.co.id is to be a site AI engines can trust enough to cite. A post built on invented statistics breaks that trust on day one.

What I still keep my hands on

Two things stay manual.

Outreach approval. When the agent drafts a guest post pitch, a podcast ask, or a directory submission, it writes the row to Notion as Pending Approval and creates a Gmail draft to me with an "rpsg-approval" subject prefix. I review, edit if needed, and approve via a slash command that re-addresses the draft to the actual prospect and sends. The agent never auto-sends to a third party in my name.

Code edits. Any change to the Astro source (schema updates, robots.txt tweaks, the site config) goes through a PR on a branch prefixed chore/rpsg-agent. I review the diff, run checks, merge.

Everything else is autopilot.

What surprised me

I expected to do roughly half the work. The actual split was closer to ninety-five five. I answered a handful of clarifying questions early ("How autonomous?" "Where does state live?" "Cron or on-demand?"), approved the initial twenty queries, and verified rpsg.co.id in Google Search Console and Bing Webmaster Tools. The rest the agent did itself.

The biggest surprise is the guardrail layer. I would not have written that on day one as a human team. I would have shipped posts, watched a stat slip through, and added validation later. The agent built it in upfront because I told it to make hallucination impossible.

What this is and is not

This is not a demo. It is a system that runs unattended, with a 6-month plan and a published scoreboard. Every Wednesday a new post lands. Every month an AI citation audit runs against twenty queries. Every twenty-eighth a KPI report drafts itself.

The next six months are trackable in public on this site.

If you are weighing whether to build something similar for your own search practice or your team's, what holds most engagements back is not the agent. It is the workflow definition. The agent only goes as far as the plan you can describe to it.

That is also the most useful outcome of an advisory engagement: turning a vague six-month wish list into a structured plan an agent or an in-house team can actually execute.

Related reading

19 May 2026 · Strategy

Train your team or hire an agency? A founder's decision framework

Capability vs capacity, the two questions that tell you whether to invest in training, advisory, or an embedded team. Founders confuse the two and pay for the wrong fix.

Read post
28 Apr 2026 · Strategy

SEO and AEO are now one practice. Stop treating them as two

Answer Engine Optimization isn't a separate discipline. It's what good SEO looks like in 2026. Here's the operating model.

Read post
Chat on WhatsApp