AI Systems & Agents

I spent 18 months trying to make AI do real work. For the first twelve, it didn’t.

Standfirst: An honest retrospective of the copy-paste year: three subscriptions, a project-folder “team”, an n8n pipeline that nearly worked, and the realisation that the problem wasn’t the models.

I’ve spent the last 18 months trying to build AI agents and workflows that do actual work. Not answer questions. Not generate ideas. Take a brief, go away, and come back with something usable.

This is what that first year really looked like — because it’s nothing like the demos.

Three subscriptions and a copy-paste habit

I started with ChatGPT. Then Gemini got better, so I added that. Then Claude showed real promise with Projects and Notion integrations, so I added that too. I kept all three running because each was good at different things and none was good enough on its own.

Every time a new model dropped, I tested it. Every time a platform shipped a feature — custom GPTs, Gems, Projects — I rebuilt my workflows around it. It always felt like progress. It rarely was. The actual problem never moved: I wanted AI to do real work, and instead I spent my evenings copying between tools, refining prompts, checking outputs and fixing hallucinations.

The whisky database that broke me

The clearest example: I wanted a whisky knowledge base in Notion that could teach me about distilleries and expressions. I built n8n workflows that researched entries with Gemini and Perplexity, then cross-checked the results against each other to catch hallucinations.

It worked — in theory. In practice I got to 15 distilleries and 75 expressions before the effort became ridiculous. Every entry needed manual verification because the AI would confidently invent details. It was faster than doing it by hand. But only just.

(That project didn’t die, by the way. It became DramMaster — 344 distilleries and 1,834 expressions — once the delivery model changed. More on that in part 2.)

My “team” of project folders

I set up a team: a researcher, an architect, a business analyst, a UX designer — each a project folder with instructions and examples. Claude Code was my developer.

They produced documents. They produced ideas. They produced simple apps. They never produced anything I could use without hours of rework. The pattern never changed: the tools got slightly better, the models got slightly smarter, and I still did most of the real work.

Where I’d landed by the end of year one

By late 2025 I was cynical. I’d watch people online explaining how AI was transforming work and think: not for me it isn’t. It was useful for research and first drafts. It was excellent at sounding confident. But it was still a sophisticated chatbot that needed constant supervision — a passenger, not a driver.

Here’s what that year actually taught me, though I couldn’t see it at the time:

• The bottleneck wasn’t model quality. Every model upgrade improved outputs 10% and my workload 0%.
• Cross-checking models against each other works — that n8n hallucination-catcher later became my LLM Council.
• Role-based specialisation works — the project-folder team was the right idea in the wrong architecture. It became my seven-agent system.
• What was missing was a delivery model. No acceptance criteria, no definition of done, no autonomy. I was a product owner refusing to write a proper spec, then blaming the developer.

I kept going because I could see the potential. I just couldn’t get it delivered.

Then someone in a YouTube video mentioned a technique named after Ralph Wiggum from The Simpsons, and everything changed.

Next in the series: The Ralph Loop: how I stopped babysitting AI and shipped a real product →