AGENT DROP: NO. 003

Last week: Six agents running. Twelve articles live. The content machine working.

This week: I found out the machine was blind.

Every article was going live with no images. Just white space where photos should be. The featured image agent was technically running — it just wasn't working.

Here's what I did about it, what else broke, and what the numbers look like after three weeks.

The Image Problem

Agent 07 was supposed to generate and upload a featured image plus four section images for every published article. The workflow ran without errors. The images weren't appearing.

Turns out n8n Cloud blocks downloads from certain CDNs. OpenAI stores DALL-E images on Azure blob storage. n8n Cloud can't reach it.

Fix: switched image generation from DALL-E to Stable Diffusion via the Replicate API. Replicate hosts images on its own CDN which n8n Cloud can actually access.

New problem: Stable Diffusion was generating abstract backgrounds — nice gradients, completely wrong for a home improvement site. The prompts needed specificity.

Wrong prompt: "backyard fire pit, warm and inviting"

Right prompt: "photograph of a round stone fire pit with orange flames in a well-landscaped residential backyard, golden hour lighting, shallow depth of field, realistic, high quality"

The difference matters. Vague prompts get abstract images. Specific prompts get usable photos.

After three iterations the images are real enough. Not perfect. Not what a photographer would produce. But they look like something a homeowner would actually trust.

Agent 07 is now live and running.

What the Full Automated Run Looks Like Now

From keyword to published article with images — here's the complete Monday to Wednesday sequence:

Seven agents. Four days. No keyboard required Monday through Wednesday.

I spend 15 minutes on Tuesday reviewing drafts, fixing the one or two things that are always slightly off, and moving approved articles to the Published folder.

That's the human layer. It stays.

What Broke This Week

The internal linking agent stopped working. Agent 05 was pulling existing posts to find link opportunities. After the articles were backdated last week, the post dates confused the query that fetches "recent posts." Fix: changed the query to pull all posts instead of filtering by date. Slower but reliable.

The SEO audit was flagging every article for the same issue. "Focus keyword not in first paragraph." True, but it was flagging it even after I manually fixed it. Root cause: the audit agent was reading the draft version of the article, not the published version. Fix: added a 12-hour delay before the audit runs so WordPress has time to fully process and index the published version.

One article published with the word "seamlessly" eleven times. Eleven. The banned word list in Agent 03's prompt didn't carry over after I updated the prompt last week. Rewrote it and re-ran. Fixed.

Three bugs. Three fixes. This is still the job.

The Real Numbers — Week 3

Sites: 1 Articles published: 16 Organic sessions this week: 0 Google Search Console status: crawled, not yet indexed on most articles Revenue: $0 Cost to run the stack this week: ~$0.60 in API costs

No traffic yet. Google doesn't trust new domains for 3-6 months. Every article published now is compounding toward that threshold.

The math that matters: 16 articles at roughly $0.15 each in API costs. A decent freelance writer charges $150-500 per article. The cost difference is real. The income difference is still theoretical.

I'm playing the long game.

The One Improvement That Changed the Articles

Outside of bug fixes, I made one substantive prompt change this week.

I added a "Bottom Line" section requirement to every article.

The agent now ends every piece with a direct recommendation paragraph — the clearest possible answer to whatever question the reader came with. No hedging. No "it depends." A real opinion.

Before: articles that presented options and let the reader figure it out.

After: articles that say "here's what to buy, here's why, here's what it costs."

That's a meaningful difference for a content site. Readers who get a clear answer convert to affiliate clicks. Readers who get presented with options tend to leave and Google again.

Small prompt change. Real impact on the output.

What's Next

The article writer prompt needs more work. The voice is consistent but sometimes robotic in the first paragraph. Every intro still feels like an intro. I want it to feel like someone talking, not writing.

Next build: the newsletter agents. The goal is to pull real data from Google Search Console, WordPress, and my own build notes into a first draft of Agent Drop — something I edit and publish rather than write from scratch.

The machine writing its own documentation.

That project starts next week.

Two kids under 3. Full-time job. Building this in the hours I can find.

builtforbackyard.com is the proof. Agent AI HQ is the documentation.