Stop Adding to Your AI. Start Deleting From It.

This week, in our own build, we deleted more than we shipped.

We permanently retired a publishing route we'd spent real engineering hours on. We abolished a fallback rule that had kept our schedule alive for weeks. We deleted dead state files and retired stale tokens. The week's most productive commits were subtractions.

Nobody writes a launch post about deletion. That's the problem. The entire AI industry has converged on the same scoreboard: how many tools your agent ships with, how many skills it bundles, how many integrations light up on the marketing slide. The number goes up, the applause gets louder, and everyone nods along as if capability count were the same thing as capability.

It isn't. Every capability you add is a liability until it proves otherwise.

The math nobody does

Each capability you bolt onto an AI assistant multiplies three surfaces at once: the failure surface (new ways to be wrong), the trust surface (new promises the user now expects you to keep), and the maintenance surface (new code paths that rot at the speed of your dependencies).

At consumer scale, you can hide all three behind averages. A million users, a broken edge capability — the metrics barely twitch. The marketing slide keeps its number. The cost is distributed so thinly that nobody feels it.

I'm building for one person: me. At N=1 there is no cohort to hide a bad capability in. Every tool that misfires, misfires on me personally. Every skill that silently degrades makes my Tuesday worse. I am the entire blast radius. That concentrates the mind wonderfully — and it turns deletion from a nice principle into a survival discipline.

This is why the last three pieces in this series were, quietly, all about subtraction. Giving my AI the right to refuse was removing the obligation to always act. Auditing it against a Tuesday was a culling process for capabilities that only worked in demos. Labeling degraded answers was deleting unearned confidence. The series looks like it's about building. It's actually about everything I'm willing to take away.

Deletion has to be scheduled, or it never happens

Addition has a natural constituency: every feature has a champion, a demo, a roadmap line. Deletion has no constituency. Nobody gets promoted for the thing they killed. So deletion, left to culture, loses — and the product accretes until the assistant is a junk drawer with a chat interface.

The fix is mechanical, not cultural. Deletion has to be a scheduled operation with the same standing as shipping:

1. A deletion backlog, next to the feature backlog. Every capability under consideration for removal gets a row, a reason, and a date. If it's not written down, it doesn't exist.
2. One subtraction per cycle. Not per quarter — per cycle. The cadence is the point. A team that deletes one thing a week develops a completely different relationship with complexity than a team that deletes one thing a year.
3. Retirement is permanent. We didn't "pause" the routes we killed this week. Paused things come back. Deleted things stay deleted, and the space they occupied stays empty until something earns it.

And the hardest rule: every surviving capability has to keep justifying its existence. Ours survive three interrogations — can it refuse, does it survive a random Tuesday, will it confess when it's degraded. Anything that fails all three doesn't get improved. It gets removed.

The real spec of a personal AI

Here's the uncomfortable corollary: if you can't name what your AI should refuse to do, you haven't finished specifying it. A capability list is only half a spec. The other half is the deletion list — the things it will not do, the routes it will not take, the answers it will not give at full confidence.

Vendors won't sell you this. "Now with fewer features" doesn't close deals. But you're not buying. You're building — for one user, who happens to be you, who will personally absorb every failure your ambition adds.

So here's my rule, running in production: for every capability I add, I owe the system a deletion. And the fastest way to make my AI better this week is the same as last week — find something it's doing, and stop it.

Stop adding to your AI. Start deleting from it. The junk drawer was never the product. The empty space is.