AGENT DROP: NO. 002

Last week: I built an agent that finds keywords automatically.

This week: I built the rest of the system.

And by Friday, Built For Backyard had 12 published articles — written, formatted, categorized, and live — without me writing a single word.

Here's what that actually took.

What I Built

Three more agents. Each one hands off to the next.

Agent 02 — Content Brief Generator Takes the top keyword from each topic category. Sends it to Claude with a detailed brief prompt. Gets back a structured brief with title, headers, word count, meta description, competitor weakness, and E-E-A-T signals. Writes the brief to a Google Doc. Appends a row to a Google Sheet.

One agent. One Monday morning. Four briefs ready before I finished my coffee.

Agent 03 — Article Writer Reads each brief from the sheet. Sends the full brief to Claude Sonnet with a 450-word prompt covering voice, structure, pillar-specific rules, and E-E-A-T requirements. Gets back a complete 2,500-word article. Writes it to a Google Doc in a designated folder.

The articles include:

Agent 04 — WordPress Publisher Reads approved articles from a Google Drive folder. Parses the content. Converts markdown to HTML. Extracts the topic category and pillar tag from embedded metadata. Creates a WordPress draft. Assigns the correct category. Assigns the correct tag. Done.

Every article goes live as a draft. I do a 15-minute review. Move it to the Approved folder. Agent 04 publishes it.

What Actually Happened

The first full run produced 4 articles on Monday morning.

I reviewed them Tuesday.

By Tuesday afternoon, builtforbackyard.com had its first real content.

Here's what surprised me: the quality was better than I expected. Not perfect. But genuinely useful. Specific measurements. Real cost estimates. A clear voice.

Here's what didn't work: every article had internal link placeholders showing as plain text. Like this literally appearing in published posts:

[INTERNAL LINK: backyard layout planning guide]

Embarrassing. Fixed it by removing the placeholder instructions from the prompt entirely.

The other issue: all 12 articles published on the same date. Looks like a content farm to Google. Fixed that by manually backdating each post to spread them across 6 weeks.

Both problems took 20 minutes to fix. Neither would have been caught without actually running the thing.

The Stack Right Now

Six agents running on a schedule:

The whole thing runs without me touching it.

Monday through Wednesday is automated. I spend 15 minutes on Tuesday reviewing drafts and moving the good ones to the Approved folder.

That's it.

The Real Numbers

Cost to run the full stack for one week:

Total weekly operating cost: roughly $0.50 in API costs.

Revenue this week: $0.

That's fine. A 3-week-old site with no backlinks doesn't rank yet. Google takes 3-6 months to start moving a new domain. The articles being published now are the ones that will generate income in month 6.

This is a delayed gratification game. The costs are immediate. The returns are not.

I'm playing it anyway.

What's on the Site

12 live articles across 6 topic categories:

Every article has:

Is it perfect? No.

Is it live, indexed, and getting better every week? Yes.

What Broke This Week

Honest section. Always will be.

The slug bug. Every post published with a broken URL like /json-title-tolowercase-replace-a-z0-9-g-replace-12/ instead of a clean slug. Root cause: a JavaScript expression in the WordPress publisher node wasn't evaluating correctly. Fix: deleted the slug field entirely. WordPress auto-generates clean slugs from the title.

Category assignment. Posts were landing in the wrong categories. Root cause: the category mapping was happening in the wrong node. Fix: moved the mapping logic upstream and embedded category data as metadata in the Google Doc itself. Now the category travels with the article.

Image download blocked. n8n Cloud restricts outbound connections to certain domains. OpenAI's image storage (Azure blob) is one of them. DALL-E images couldn't be downloaded after generation. Fix: switched to Replicate API for image generation. Replicate hosts images on its own CDN which n8n Cloud can access.

Three real bugs. Three real fixes. That's the job.

What's Next

Agent 07 — Image Agent is live. Every published article now gets a featured image and 4 interior section images generated automatically via Stable Diffusion on Replicate. I’m nearly done troubleshooting.

Next session: improving the article writer prompt. Better SEO signals. Better voice variation. Focus keyword in the first paragraph.

Then: the newsletter agents. The goal is to automate the first draft of Agent Drop itself — pulling real data from Search Console, WordPress, and my build notes into a draft that I edit and send.

The machine writing its own documentation.

That's where this is going.

I'm building this nights and weekends — between a full-time job and two kids under 3 — to see how far AI agents can actually go in the real world.

builtforbackyard.com is the proof of concept.

Agent AI HQ is the documentation.