We Deleted Sixty Percent of Our Own Instructions

In August we cut about sixty percent of the instructions we give the AI tools we use every day.

The drafts that came back afterward were sharper, and they arrived finished.

Here is what came out, and the test we use now.

One: the same rule, written twice

Our banned language list lived in two files, a global instruction file and a project knowledge base. Written at different times, the two had drifted. One banned five words outright. The other filed the same five words under flag this and ask a human.

Hand a model two documents that disagree and it does the sensible thing. It stops and asks.

We kept one copy. The other file now holds a single pointer to it. The questions stopped.

Two: decisions filed as questions

We had a middle tier called surface it and let a human decide. Every word in it was a decision already made.

We moved each one into the rule, once. A banned word now gets the line rewritten before the draft reaches us.

Three: the route

The biggest cut was the step-by-step. Do this, then this, then check that. Pages of it.

Work written against a route is capped at how good the route was on the night somebody wrote it.

We kept the standard instead. What finished looks like. What it must never do. The things only we would know. Getting there is the model's job.

What stayed

The rules we kept describe taste, and those carry the work. Steps and reassurances mostly told the tool to be careful, and careful is what it gave back.

A test you can run this week

Leave your setup alone for two weeks. Then open it and read it cold.

If you cannot remember what half of it was for, neither can the model. That half is the cut.

Keep the standards. Delete the choreography.

A second pair of eyes

This is most of what we do in a one to one engagement. Somebody arrives with a set of tools and a set of rules, and the work is deciding what stays. You or your small team, us, and a system you actually run at the end of it. The word we use for that system is a harness.

Tell us what you are working with: https://thehouseofcuriosity.ai/consulting-services