How I’m Learning to Put Down My Keyboard and Trust in AI

This week I really hit a wall with the Profit Swarm challenge – I almost gave up yesterday. This post is about how I am breaking the code-everything habit, (and also a quick update on AI Directory Builder-it’s made its first directory!)

Learning to Software Engineer in the AI Era

Crying onto my Keyboard

I looked up from my code editor and shook my head… It had been 2 solid days of coding the AI Directory Builder agent (build site part). I felt like I had failed you.

This challenge was supposed to be all about AI automating, but I’d let myself slip into old (engineering) patterns:

• Hard-coding automations
• Spent too long making & optimising HTML templates
• Ended up as one giant function

The two decades I’ve been writing code have left me in a solid pattern: Just write the code.

I thought it’d be easier to let go; to break the habit. 

…Mind you I was using a lot of AI to assist me:

• Lovable for roughing out templates
• Cursor AI with variety of models (Gemini Flash, Claude Sonnet Max)
• AI calls in the code to improve content, choose emoji’s etc.

But it still took a lot of dev hours to achieve this. As I thought about it I wondered:

Is it that I chose to write the code more often than ask the AI agent to do it?

Would it be much quicker if I’d scoped it cleanly and walked Cursor AI through it?

Methodology of AI Biz Experiment #1

I achieved 75% automation of my first agent (AI Directory Prospector) using mostly ‘off the shelf’ tools: n8n/Gumloop/Relevance AI & RapidAPI APIs. This seemed promising, but I hit lots of in-software bugs that I couldn’t quickly work around.

I told myself ‘I’ll just code the second agent to compare the two approaches’.

I still think it was a good thing to do, but I wish I’d checked in with myself more often – the aim was to automate everything here, and although I can write-off some setup time, it feels disingenuous to call it ‘all AI’ with that much time spent.

• Agent 1, achieved 75%~ success with off-the-shelf workflow builders, mostly prompt engineering; but I hit issues, so..
• Agent 2, fell deep into writing logic, but then didn’t hold myself fully accountable. Stopped myself 1/2 week deep into logic. After the fact, 78% automated.

What do you think? Was this a useful step – letting myself write the whole agent in code (leaning on AI for some of it). I think the outcome will be more stable than existing workflow tools, but can I call it automated if it takes a lot of setup?

What’s Working so Far?

Now that we’re 50% of the way through experiment 1, I want to share what I think works well, and will carry on working well as the AI storm develops:

1. Your Business as an API – I continue to believe this is a crucial part of how AI business is going to play out. It’s working well for me
   1. Modular Logic – Having everything broken down into reusable endpoints
   2. Gatekeep – Having your biz logic behind an endpoint ensures IP
   3. Middleman other API’s – Keep all your credentials internal & throttle usage
2. Some Workflow Layer – I think until the end of this year we’ll still have some form of a workflow layer. Whether this is internally written, or it’s n8n/Gumloop/Relevance AI is to be seen
   1. Abstracted from API – An overseer agent which runs regardless of failures
   2. Workflow design – Orchestrating workflows via UI

I think all of us who are embracing the AI evolution need to be seriously investing in 1. (Business API), and seeing 2. (Workflow layer) as interchangeable.

Fundamentally the IP of future businesses will be kept in their APIs, and we’ll probably all move Workflow tools as and when each improves/provides better value.

What’s Not Working?

Me!

I think the thing that’s slowing this experiment down this week is me. I keep getting lost down familiar but less-than-productive code holes.

I always knew this challenge would be a mixture of shiny new tech and brutal adaptation, but I hadn’t counted on the need for such an ego death so early on!

• I need to keep reminding myself that our AI future is really close, and re-tool myself with an AI layer between my keyboard and my IDE
  • Have discipline making PRDs and using AI first
  • Check myself if I’ve been coding something for an hour or more
• Workflow tools I’ve tested so far don’t quite offer the right set of tooling

It’s a fundamental mindset shift; to embrace AI fully as the maker.

Finding the Right AI Blend

I do believe that currently it takes a blend of approaches to fully AI automate most tasks.

I do believe it’s good to experiment.

But I want to dream bigger.

I want to make this MORE automated.

In the end, I’m taking this on–going through this suffering so you don’t have to. 

So let’s take a minute to project ourselves into the future 1 year from now, what will AI automations look like then?

My AI Automation Predictions for 2026

The more I dig into AI automation, the more I recognise the philosophy change that is going to be needed. That, I think, will be the only thing which slows progress here.

I think in 2026 we’ll automate like this:

1. Either invent/specify the workflow, or let AI observe it
2. AI will compute a PRD/FlowSpec document
3. AI will go away and build tools it needs (Business API Endpoints)
4. AI will ask/retrieve any access it needs to external services
5. AI will build out a ‘workflow agent’ (either in a tool like n8n, or custom)
6. AI will run & test the whole flow
7. AI will report:
   1. “I’ve made the flow for cold-outreach, it has 10 steps.”
   2. “I’ve tested the flow and it should run fine, it will take approximately 32 seconds and consume $0.42 in AI tokens per run.”
   3. “This workflow consumes services: RapidAPI, X API, Sendgrid.”
   4. “I made 6 new tools (enrich_data,linkedin_api_retrieve,format_email,generate_ai_video,send_cold_outreach,add_to_crm), and reused 4 existing tools (…)”
   5. “Do you want me to make this live, run it 3 times a day at the optimum send time, then we can scale it up to 300 times a day over the next 8 weeks?”

For some tasks perhaps the AI will falter and need some engineer oversight.

For most? It’ll just work.

Would it be useful to you if I built this as a tool? Comment Below if so.

I thoroughly believe this is how the system will look in 2026, and it will be a wild time with millions of humans writing these things and setting them live.

Mindset Shift; a New Paradigm.

We, as makers, are going to need this fundamental mindset shift. Our making is going to change, and fighting it is only going to hurt us more in the end.

With that in mind, I believe we should lean heavily into preparing ourselves for the oncoming tooling. I intend to keep working on my Business API, (making myself do PRD’s properly and writing annotated clean code primarily directed by an AI agent).

I will also keep developing FlowSpec, because I think having our workflows already designed in a machine-readable way will speed up the process. (Appreciate any input on that from you!)

I’ll keep tweaking my blend, and sharing the challenges, and anything else I can which I think will be useful to you.

My New Approach to Building with AI

No Woody, you can’t ‘just build it’.

1. Design workflow with AI assistance
   1. Produce a PRD, a FlowSpec, and from that flow chart
2. Split everything out into micro-tools/modules
   1. Cursor build (for now) each micro-tool/module
   2. Later give AI descriptions/access (via api/local)
3. Get AI to stack them into workflows
   1. Continue testing workflow tools
   2. Make a lean tool myself if none work without issues (with AI 😂)

As ever, really appreciate your input on this, comment below if you can see something I can’t!

Now, want to see my first AI-built (mostly) directory?…

AI Directory Builder Update

First AI-built Directory Live!

Today I fired the API command to make vibegames.io – the first AI generated directory by my AI Directory Builder agent (it took a 6 minute API call to build!)

I’m fairly happy with the output here, the directory looks clean, everything works, and I don’t even mind the AI layout that much.

VibeGames.io is:

• Generated in pure HTML for fast loading speeds
• SEO optimised
• Got working submit form (defended by Cloudflare Turnstile)
• Got the basics (privacy policy, robots.txt, sitemap, about page)

But I’m less happy with how automated the process was.

If we take a look at the AI Directory Builder agent workflow:

… we can see that we’re now  up to step 5 on the flow:

1. Register domain – API endpoint, AI Automated
2. Setup hosting – Currently manual
3. Retrieve starting data – 75% AI Automated, though it needed more guidance than I’d have liked
4. Build directory files – This took a lot of code (this is where I got lost this week), but does now fire from an API – 100% AI Automated
5. Deploy & Test – Manual for now (will be quick to automate)

… so all in I’d say on its next run it’ll be 75% automated. 

I guess that’s an achievement, but I still feel like I need to find a way to automate more of it. 

---

Watch this space for future updates on the whole AI Directory experiment (I always update the experiment page first). 

I’m going to try to fully complete the first 2 agents and update you, then I’ll move onto the AI Marketing Agent (which I know a number of you are waiting for, thanks for bearing with me!)