It was close to 1 AM. I looked at what I had built over the previous three weeks and realized I had made myself the bottleneck on every sentence of my own creative universe.
Per-quote canon rulings. Individual approval on each passage before it could be committed as locked lore. I was three days into reviewing quotes that were already sitting inside books I had already written.
I had built a protection layer that was protecting me from finishing.
i.How It Starts
AI-assisted creative work has a real failure mode and it is not the one people write about. The failure mode everyone warns you about is the AI going off-script, inventing lore, drifting from the canon you have established. That is a real risk. I have seen it.
So you build gates. You add checks. You create a review process that keeps the creative work tethered to what you actually decided. Each gate is rational. Each one addresses something that went wrong or could go wrong.
The problem is that gates compound. Three weeks of reasonable, individually justified checks had produced a process where I was the decision-maker on every atomic unit of lore, in real time, on demand. The system that was supposed to protect my creative work had made forward motion contingent on my uninterrupted availability.
That is not protection. That is a different problem wearing the same clothes.
ii.The Audit at 1 AM
The conversation that changed it was not a planning session. It was a moment of honesty. I looked at the review queue and named what I was seeing: I had become the mechanical step in a process that was supposed to be protecting my creative judgment, not consuming it.
The per-quote ruling model had made sense when the goal was airtight canon. But I was not defending against canon drift. I was approving things I had already decided. The quotes were already in the books. The decisions were already made. I was reviewing my own output, one sentence at a time, to gatekeep a universe I already owned.
The overhead had outrun the risk it was managing.
The system you build to prevent your AI's mistakes can become the mistake. Gates compound silently, and no single gate looks like the problem.
iii.The Fix Was Altitude, Not Removal
The answer was not to remove the protections. The creative universe still needs canon discipline. The AI still needs to work within what was actually decided, not what it infers from adjacent material.
The answer was to move the gate to the right altitude.
Instead of per-quote approval, the model shifted to phase blocks. A block of work runs through a full adversary review, an AI evaluator playing the role of a critic who knows the canon and is looking for drift. Then I approve or refuse the whole block. One decision gates one block, not one decision per sentence.
The protection is still there. The adversary is still there. My judgment is still in the loop. But it is in the loop at the level where my judgment actually matters, at the block, at the arc, at the thing that has shape and consequence. Not at the sentence that could be checked by a deterministic rule.
The creative overhead dropped. The forward motion returned.
iv.What This Cost
Three weeks. That is the honest accounting of how long the wrong-altitude gate ran before I caught it. The work was not wasted. The lore is real. The discipline that motivated the gates is the right instinct.
But three weeks of available time in a creative project with real stakes is not nothing. The manuscript I have been building for years has a series waiting behind it. Time spent reviewing already-decided quotes at 1 AM is time that did not go toward finishing.
The lesson I keep returning to: no single gate looks like the problem. Each one is justified on its own terms. The accumulation is what you have to audit, and accumulation is hard to see when you are inside it. The signal that something is wrong is not usually “this gate is bad.” It is “I am the bottleneck on my own work, and I do not know exactly how that happened.”
When that feeling lands, the question is altitude. What are you gating? At what level? Does the gate sit where your judgment actually changes outcomes? Or has it slid down to the level where it is consuming judgment without adding it?
Build the protection. Audit the altitude. Then build it again at the right level.
Drafted with Bishop, my AI partner.
Words picked, edited, and approved by me.
Model provenance: Claude Code (Claude Opus and Sonnet)