A Perfectionist's Permission to Ship

I had a backlog of ideas and a field moving faster than I could publish, and I kept not publishing. The usual story is that drafting is the bottleneck and AI fixes it. It did. I can draft faster than I ever could. And I still wasn’t shipping, because drafting was never my bottleneck. Editing was.

Here is the trap a perfectionist hits with these tools. The draft now comes out in a tenth of the time, and then you spend the same three hours you always spent polishing it, because “close enough” is not in your vocabulary. The tool made the cheap part cheaper and left the expensive part exactly as expensive. So the output barely moves, and the backlog quietly goes stale while you make three things perfect instead of shipping ten.

The fix was not more discipline. I have plenty of discipline; the discipline is the problem. The fix was a system that decides, before I start, how much editing a given piece is even allowed to get. I tie it to how the piece is attributed, on a scale I already use for how much of it was AI.

An AI-E piece, where the value is the argument and the AI helped me get it down, gets one pass. The only gate is “is the argument mine, and is every claim defensible?” If yes, it ships, even when I can see five ways to make the prose shinier, because making it shinier is against the rules for this tier. An AI-C piece, where the structure and framing are the work, gets two or three passes against a harder gate: “could I defend every paragraph live?” An AI-A piece, the anchor kind, gets the full treatment, because the gate there is “this is my best thinking on the subject” and that is worth the time. The level is a budget for my own perfectionism, set in advance, when I am calm, instead of at midnight when every sentence looks improvable.

The move that makes it actually work is publishing the rules. Once the gate is written down in public, continuing to buff an AI-E post past its gate is not diligence. It is a violation of my own stated method. The rules become the discipline, so I do not have to supply it sentence by sentence. (This is the same reason you write a convention down instead of trusting yourself to remember it in the moment. The version in your head loses every argument with your own perfectionism.)

None of this is new, exactly. It is code review. Nobody reviews a CI config the way they review a core algorithm; you match the depth of review to the blast radius of the thing. A throwaway enhancement and a load-bearing argument should not get the same editorial pass, and pretending they should is most of how a backlog dies.

The gate also tells me what to do with the drafts already rotting in that backlog. A perfectionist sits on a piece because it feels stale, but “stale” turns out to be two different conditions. If the hook expired, if it led with a tool version or a news cycle or “I just tried X,” then yes, the post is dated. But if only the hook expired and the argument underneath is a reframe or a named idea, the post is not dated at all. “Infrastructure over policy” does not have a shelf life. So before I kill an old draft, I check which part actually died. If it is the hook, that is twenty minutes rewriting an opening, not a reason to throw away the thinking. Most of my “too old to publish” backlog was a stale first paragraph wrapped around a fine argument.

So this is infrastructure, not willpower. I did not get better at letting go of a piece. I built a thing that lets go for me, and then I published it so I could not quietly take the permission back.

PS — the irony that I nearly over-engineered the system for not over-engineering my writing is not lost on me. I shipped it at AI-C and moved on, which is the whole idea.