Everything is possible.
That is the problem.

The cost moved.

A part is nearly free to add now. You still pay for it on every turn after. Add a few and watch the cream hand fall behind the brass one, which is true time.

Two hundred thousand tokens before my first word.

That is what my setup loaded on every turn: rules, hooks, routers, context. My million-token window evaporated like water.

Ask one thing: did it change a decision?

I tested every part of my setup that way. If the answer is no, it comes out.

Easy to add. Hard to take off.

Each part clicks on in a second. Nothing clicks off. To remove one, you have to lift the pawl on purpose.

What works grew out of something that worked.

John Gall wrote that down in 1975. Start with a simple thing that keeps time, then grow it one part at a time. Slide it out and find where it stops.

Three ways to know you over-built.

• One caller.A layer used in one place.
• Might need it later.Build it when you need it, not when you foresee it.
• A system watching a system.My optimizer only saw the first 3,000 characters of what it tuned.

One rule doing real work beats ten that sound wise.

When I retired the routing layer, my instruction file went from 1,013 words to 734. Pull the slider and watch the hand steady.

Before you add, name what you will delete.

Pick the part you will give up first. Only then can you add one. The count holds, and so does the drift.

Chuck it. It’s extra weight.

One part, one shadow. It keeps time. The simplest thing that works.