A Tale of Two AI Agents (Partly) Building Web Directories

Maybe we’re getting somewhere. This week I share the output of the first half of this experiment: a shiny new web directory prospected, made, and deployed in 33 minutes – 90% of the work done by AI agents or automations.

AI Made Web Directory Build

First (Partly) AI Built Directory

It was so easy.

Not.

In the last week I’ve been refining the first two agents which prospect and then build a niche directory website from scratch.

The first refined directory is live, you can check it out here:

(Previously I’ve shared another, vibegames.io, which was also made by these bots, but it wasn’t a true test as I walked it all through manually.)

As you can see the site is fairly typical directory fare.

How it Went (First 2 Agents: AI Directory Builder)

The improvements I added helped this go smoother than it might have, but overall I am still not happy with how much human input it took.

It’s cool to make a brand new web directory from idea to shipped site in 33 minutes.

It’s not cool that it still fell over a few times.

If directory businesses were all about initial throughput, e.g. creating lots of directories, I’d slim this down even more; further refining the flow to reduce human intervention.

But it works, I guess 🤷‍♂️.

Improvements & Changes Added before I Started:

I added these bits based on my notes:

• Added screenshot maker API to generate screenshots for each directory object
• Switched out image generation for GPT Image 1
• Google Gemini Image Generation is now blocked in UK… had to sidestep OG sharing meta image generation
• Made a new back-end management page: View projects-in-hand
• Passed images through Kraken

AI-Human Handover

I had to do too much to call this “AI Automation” in a complete sense. Compare this to doing it the oldschool scripted way though – I definitely at least halved the labour element.

Here’s what AI left me to do:

• Choose the Logo
• Approve the domain choice
• Setup the hosting, email, SSL certificate
• Rebuild starting data
• Tweak some images
• Set up Turnstile, Google Analytics, etc.

Refinements to work in Later

Here’s what I’d still change before I’d run this 10+ more times:

• Nudge the AI to find more .com options for domains, .io doms cost $68 😬
• Find a more automatable hosting setup, (plus SSL costs £30 with my current)
• Make the AI also find affiliate application links for me if I list any tools
• Automatically fly images through Kraken for lossless optimisation
• Initial data generation routine sucks – o3 Chat GPT annihilated it, need to agentify it
• Add more stylistic variety / properly leverage CSS palette

Overall % AI Automated: 90%

AI / Human Split:

It’s hard to quantify specifically how much time this automation has saved me, but I estimate scripting & AI have taken 90%+ of the work out of the process. I figure actual AI model work made up about 60% of all work.

• AI & automations completed 90% of the work
• “AI” probably did 60% of the work
• AI & automations ran 52% of the work in pure time spent (17 minutes runtime vs 16 minutes me time)
• Total directory deployment time, from scratch = 33 minutes

How I Feel Now (About these 2 Agents)

I’ve learned a lot.

I’ve built up a lot of reusable code/workflows/agents/prompts.

It’s taken longer than I expected though, and it doesn’t feel super ‘AI’ leveraged yet; but we are working on the coal face here. In terms of the AI business challenge, this feels short of the mark. That just makes me want to try wilder approaches to share with you in the coming months!

What’s Next: AI Directory Maker & ProfitSwarm

I’m going to finish out this experiment then go back to the drawing board for more aggressive approaches & to streamline what I have set up and learned. Somewhere in there I am confident I’ll find some leverage which’ll turn some profit.

Next up: I’m going to build a AI Start-up Marketing Agent to launch a suite of directories made this way.

How I can Help You with ProfitSwarm

Any questions?

Anything you think would help with my challenge?

Anything else you’re keen for me to experiment with?

😎 Please comment below!