My Wise Buddy and I #2: Deciding Less, Twice

My Wise Buddy and I #2: Deciding Less, Twice

I have an ordinary office job in the food industry — mostly planning work, no engineering — and I build this AI project in the hours around it. I watched AI from the stands for three years and only opened a terminal for the first time this year. This is the week I found out that my main contribution to my own project had quietly become: being the bottleneck.

The sentence I had been reading wrong

“The human is the bottleneck.” I’d read that line many times and, honestly, took it as a complaint shaped like a compliment. Look how much depends on me.

Then I looked at where my hours actually went. Reading long AI answers closely. Then either approving something I didn’t fully understand, or asking a follow-up question that produced another long answer to read closely. I wasn’t steering anything. I was proofreading.

The thing that clicked: “the human is the bottleneck” is a diagnosis, not a prescription. Removing myself isn’t the fix, and neither is reading faster. The fix is to split decisions into two piles.

Pile 1 — only I can answer this. What problem are we solving. What is the goal. What counts as success, and how do we verify it. Who reviews. Where is the irreversible boundary — the thing that must never happen without me. These deserve dense, slow conversation, and they deserve it early.

Pile 2 — everything derived from pile 1. Usually in a domain I don’t know, and usually with no meaningful effect on the outcome. The AI decides, proceeds, and records what it decided.

Naming the two piles took ten minutes. The hard part is that the border between them drifts — always toward pile 1, because asking me feels safe to the model and feels like control to me. So the border needed its own supervision. Two reviewers:

1. An autonomy gate. Given the plan as it stands right now, is it safe to run without a human in the loop? Pass, and the autonomous stretch begins.
2. A conversation-depth check. In the conversation stretch, is the AI asking me things it could have decided itself, or looked up in a file? Is it escalating detail that doesn’t need me?

I turned this into a global rule the same day. Its shape:

• Two axes for every decision point: risk (if this is wrong, is the final effect irreversible?) and authority/information (does this need something only I know, or something only I can authorize?).
• Four routes out of that judgment: AI decides / verify against source material / ask the human / safe stop. The fourth is the one I keep forgetting exists — low confidence plus high impact means stop and report, not guess and proceed.
• Two phases: an agreement phase, then an autonomous phase, with the gate between them.
• A calibration loop, which is the part I’m least likely to have gotten right on my own, because it doesn’t trust my judgment about my judgment. The border isn’t set by feel; it’s moved by two measured rates. Autonomous-reversal rate: how often I overturn decisions the AI made alone — high means pull that area back into conversation. Unnecessary-question rate: how often it asked me something it could have resolved — high means push that area toward autonomy.

A second idea fell out of the same morning: three notes per project. A fact note (what is true, with sources). A decision note (why we chose this, and — the field that matters — who decided: me, or the AI alone, plus a later verdict: held, overturned, discarded). A prompt note (the history of the conversation). Counting “decisions the AI made” tells you nothing useful. Decisions the AI made × later overturned gives you a concrete signal about where to review or narrow autonomy.

Then I built the whole thing twice

Here’s the uncomfortable part. That rule was written by the session I’d been working in all morning — the one that had watched every mess that produced the idea. Which is precisely the session you should not let grade its own homework.

So that afternoon I handed the same prompt to a fresh top-tier session with no context at all. Clean room. Then I put the two versions side by side and merged them.

I expected one to win. Instead, reading both, what I found split by layer.

In this comparison, the context-holding session was stronger on the core. The precision of the gate’s pass conditions. The data model — including that decision-maker taxonomy. The hygiene rules about what not to copy where. It reads like it was written by someone who had been burned in specific places, because it was.

The clean-room version added instrumentation, enforcement, and automation — four things the other version simply didn’t have. A written contract that makes each gate judgment auditable as a document diff. The quantitative calibration loop (those two rates). A JSONL export turning the decision log into training material. A reminder hook that fires right before the AI asks me a question.

Look at that list again. The clean room built the machinery that measures whether the rule works. The session that invented the rule couldn’t see that gap — it was the rule. Skin in the game buys precision and charges you perspective.

Final version: the context session’s core, plus the clean room’s instrumentation layer.

If you’d rather read the thing than my description of it, the merged pack is public: BehindTheBuild — the two rules, the gate checklist, the contract template, the reminder hook, and appendix/COMPARISON.md, which is the actual scorecard between the two versions. Fair warning: I work in Korean, so the pack documents are Korean; the README summarizes each piece in English.

Was doing it twice a waste?

Cost: the same work twice, two contexts, one afternoon. Return: a hypothesis about which layer to trust from which condition — one I can now test on the next rule instead of guessing. So it’s a working rule of thumb, not a law:

• Draft the core in the session that lived through the problem.
• Draft the measurement and enforcement in a session that has never seen it.
• Merge them. Don’t pick one.

None of this is a new idea. Separating the author from the reviewer is something every team already knows. What surprised me is that it applies to the tool, not just to people. I had assumed more context is strictly better. It isn’t. Context is a lens — sharper in the middle, curved at the edges.

And the joke writes itself: I set out to make fewer decisions, and the immediate cost was two extra decisions about how to make fewer decisions. I’d take that trade again.