How I start an AI-driven build project

Last week I gave a first hint about playitsmart.nl on LinkedIn. Since then I keep getting the same question in different forms: how do you actually start? Open Cursor, type a few prompts, and see how far you get?

No. Not at all.

Before I let a single line of code be written, I spend days on something far less exciting: writing, thinking, setting up structures. It doesn't look like building. It is fundamentally the most important part.

The most important lesson: context, context, context

If there's one thing I've learned working with AI, it's this: make sure the system always knows the context. Of every choice, every document, every design principle. Context when you're building, context when you're reviewing, context when you want to change something.

AI without context is a smart intern who doesn't know which company you run yet. AI with context is a colleague who thinks at your level, in your direction, with your boundaries.

The entire foundation I describe below revolves around that one point. Capturing context, making it accessible, and keeping it alive.

Step 1: get thoughts on paper

I started with a single document containing my raw thoughts. What I want to build, roughly how, and why. No more than two pages, free writing, no polishing.

Then I had a conversation with Claude Opus. Not to build anything, but to test my ideas. Questions like:

"I don't want to be a pure day trader. It needs to be based on fundamental analysis, with momentum included. What could that look like?"

After a few iterations, Claude had drafted a trading design document. Not perfect, not final, but a solid starting point. I did the same process for the architecture. What do I want to publish, in what languages, what data, what integrations, what hosting.

Ten iterations later I had two documents describing the entire project. Trading system design and technical architecture.

Step 2: create a Claude Project

With those documents in hand, I created a new Claude Project. There I capture how I want to work:

• How we communicate (Dutch, no em-dashes, trading terms stay English)
• Who does what: I think and review, Claude thinks along and designs, Cursor builds
• Which documents Claude should always have at hand

This is not decorative. It ensures Claude knows where we are at the start of every new session, what we decided yesterday, and which rules we follow. No repetition, no lost context.

Step 3: living documents

From the start I work with six essential files:

• NOTES.md: current focus, blockers, status
• TODO.md: active tasks per week
• DECISIONS.md: every substantial decision with rationale
• LESSONS_LEARNED.md: problems we encountered and how we solved them
• WORKFLOW.md: how we collaborate
• ARCHITECTURE.md: technical quick reference

These documents are not static. They live along with the project. I ask Claude to update them automatically based on what we decide and what we encounter. At the start of every new session I say "read the documents" and he knows where we are again.

In my experience, decisions is the most important. Every choice I might forget or doubt later, I capture with date, context, options I considered, and why I chose. Months later I don't have to think again. It's there.

Lessons learned is my second favorite. Every time I run into something, I capture what the cause was and how I solved it. AI sometimes makes the same mistakes again. With a lessons learned file you prevent that pattern.

Step 4: the six-eyes principle

My way of working is deliberately not "I tell Cursor something and he builds". That's two eyes. Too much risk of mistakes or going off track.

I work with six eyes:

1. My eyes: I think out what I want, I review the output
2. Claude's eyes: he thinks along, designs, writes the prompt for Cursor, reviews afterwards
3. Cursor's eyes: he builds based on Claude's prompt

Concretely this means:

• I ask Claude to think something through
• Claude provides a structured Cursor prompt
• I paste the prompt into Cursor
• Cursor builds, I review the output
• I paste the result back into Claude
• Claude reviews and suggests improvements

This is slower than two eyes. It is far more reliable. Three different perspectives on the same piece of work.

Step 5: account setup parallel, but structured

Alongside the document work I arranged the practical things. IBKR account application (which takes 3 to 10 working days, so that needs to start early). Choosing a data provider, in my case Financial Modeling Prep Ultimate at 100 dollars per month. Domain, hosting, email.

This didn't go smoothly. On day two and three I was doing four things at once: configuring Vercel, setting up Render, configuring Git, arranging email. I was constantly switching, and that went wrong. My local Git wasn't configured properly. My shell environment was working against my Mac.

Lesson: even though AI makes it possible to go fast, working calmly one thing at a time is always faster than tackling everything at once. You feel that it can go fast, so you want to go even faster. That's the trap.

The sum total

By the end of phase zero I had:

• Two complete design documents
• A Claude Project with instructions
• Six living documents
• A way of working (the six-eyes principle)
• Five accounts arranged or in application

Not a single line of code. But a foundation on which Cursor can work productively without me having to think up every prompt again. Without having to explain every decision again. Without mistakes being repeated.

My recommendation

If you're starting with AI build: in the first week, invest mainly in the foundation. Documents, Claude Project, way of working, living context. No sexy LinkedIn screenshots, but the difference between a project that succeeds and a project that gets stuck in chaos after three weeks.

And remember: context, context, context. That's the whole trick.