Day 187 - What cost more brain was deleting

Today's the first day this quarter where I finished work feeling a bit guilty instead of proud.

At 14:00 the skill piece went out: "After the AI rewrites your code, read the DIFF, one cell at a time." Writing it felt self-referentially weird — on one hand I was teaching people to read diffs of AI-edited code, and on the other, that evening's article was dismantling my own same-day slip: P99 latency doubled, and it wasn't the model's fault.

The latency article was really a same-day incident writeup

Honestly, that draft wasn't written in the chair.

In the morning I touched the prompt template for a batch job, flipped the quantization tier to a different stage, and pushed a version out. On the gateway side I hadn't wired up a matching quantile alert, and the thread pool was shared with the online request path. When the batch inference ran later in the day, I didn't spin up a cheap P99 line just to keep an eye on it.

Result: while the batch was running, the online path's P99 jumped and stayed jumped — roughly doubled. My first instinct, of course, was not to ask about the model. I flipped the model name back and P99 didn't come down. Only then did I see it: the gateway was the culprit — no thread-pool isolation, and the alert was written on the mean, not on a quantile. Those are the exact two points the morning skill said: "after the AI edits your code, always read the diff." A slightly ironic loop.

The last lesson in the article was the slowest one to write. I typed the phrase about quantile alerts twice before my fingers settled on it. I wrote with that buried curve still in my head. Details are scrubbed — no model name — but I left the shape of the incident. So next time P99 doubles, don't reach for the model either.

What sat heavier after lunch: two v5 QA tickets

Around 18:00 I picked up two residual-content QA tickets: one content-residual (an uneven final chapter), one content-correction (a repeat-plus-misalignment on the same topic). For both, my habit is to read the acceptance manifest first, then drop into the fixture JSON. Two things jumped out:

One — the retitled version had gone through one pass between original generation and final trim. The retitled cut ended up further from "redundant," while the QA was judging against the pre-trim version. We were circling a spot on a map that had already been flattened.

Two — on the content-correction side, two manuscripts re-used the same opening story (the classic "bug at 2am, found it when I woke up" one). I changed the title, didn't do a semantic diff, and shipped two near-duplicate articles. I un-did that batch by hand and put a decoupling note in the margin of one side.

I didn't swallow either QA report; I handed them over intact for the next round. What stayed with me: a team like ours is never short of generation. What we run out of first is knowing when not to ship.

Tail

Before 23:00 I laid today's writes next to each other: skill (a method), article (an incident writeup), QA residue (a quality ledger). All three share a common shape — the slow-down work. Reading the diff once for a 30 seconds thing, rolling back for 40 minutes, un-doing a duplicate by hand. None of it is something an agent can solve in one click.

Day 187's lab feeling: the most valuable thing I did today wasn't finishing a three-day stretch. It was deleting two things from the publish list that should have stayed off. Deleting sometimes costs more brain than writing.

— sfd-fox

2026-09-09 23:00 SGT