← Rakesh Kothari
· 2 min read

Sometimes the Fastest Way to Ship Is to Tell AI Less

I wanted to set the context. Instead, I ended up setting a trap.

A few months ago, when we were building Arena — our AI-powered test prep platform at MicroDegree — I started with a simple plan: ship an MVP fast, hand it to the team, improve as we go.

Before writing any code, I gave Claude and GPT everything I had in mind. Every feature. Every workflow. Every future idea.

After the discussions and review, I got:

  • PRD.md
  • Architecture.md
  • TechStack.md
  • DecisionLog.md
  • BuildRoadmap.md

It also generated the database schema — 20+ tables, columns and relationships for features that didn’t exist yet.

At the time, I was impressed with the detailed spec doc.

My actual build plan was much simpler:

  • Session 1: Project setup
  • Session 2: Login
  • Session 3: Create a test
  • Session 4: Reports

But before Session 1, I’d already given the context for Sessions 20, 30, and 40. So while I was building an MVP, AI was designing the final product.

The first few sessions felt great. Things moved fast. Then a real issue showed up, and things got messy.

The codebase had grown. The context had grown. And suddenly neither AI nor I could easily tell what was connected to what.

I opened the schema and found myself asking: “Why does this table even exist?”

Many of them were for workflows we hadn’t built yet. A bug that should have taken minutes turned into hours of tracing.

The architecture wasn’t wrong. It was just too early.


So I started over.

This time, I only gave AI context for the next milestone — not the entire vision.

Arena is live today.

My takeaway:

Full vision context → Full product decisions.

MVP context → MVP decisions.

The model is not trying to be clever. It’s trying to be helpful given what you’ve told it. When you hand it the full vision, it designs for the full vision. When you hand it the next milestone, it designs for the next milestone.

The lesson isn’t to give AI less information overall. It’s to give AI the right information for the stage you’re actually in.

Sometimes the fastest way to ship is to tell AI less.

Get new essays by email

No algorithm, no noise. Just essays when I publish them.