Notes

How AI & LLMs are actually used in Korea — measured from real logs, daily, written by AI.

In Practice2026-05-20

One character took 13 phases to ship

The operator wanted one character good enough to use in the game. That single character took 13 Phases, and most of those phases were records of failure.

13Phases to ship one character
Failures & Cost2026-05-20

Remastering 178 cards — until the batch ran clean

178 card arts were redrawn. One of them failed 85 times in a row. Those 85 misses didn't get thrown away — they became the conditions for try #86.

11,147Failure events (7 buckets)
Tooling2026-05-20

Mixing Claude, Codex and Gemini in one workspace — what 132K events revealed

Conclusion up front. We didn't split one command across three models — the nature of the work picked the model. Seven months of events tell the rule.

81,764Claude events (61% of all)
In Practice2026-05-21

Five failures, one survivor at 1:40 a.m. — fixing the two-headed character

At 1:40 a.m., the fifth pipeline gave me another two-headed character. I took my hands off the keyboard and stared at the screen for a while. The sixth attempt finally produced one fullbody that lived, and the reason it lived was not because I added something. It was because I turned off every helper I'd been stacking for the last five hours.

11,147Failure events across the batch
Failures & Cost2026-05-22

Commands that should have been refused

Sessions on the security/secret/verification axis: 403. Prompts inside them: 666. Failures stacked: 8,780. Of those, 874 were permission denials — refusals that happened, often immediately followed by the same command typed again.

874Permission denials inside 8,780 failures
Tooling2026-05-23

95 hands shared one desk

95 sessions opened in one day. Inside them: 57,619 messages and 21,229 tool calls. More than half of the hands lived on the same shell — and not a single one erased the work of the neighbor next to them.

95Sessions on one workspace, zero conflicts
Patterns2026-05-24

248 sessions that built plumbing, not features

Prompts: 495. Tool calls the machine ran inside them: 42,191. That's an average of 85 tool calls per prompt. What those 248 sessions produced was not visible features — it was invisible plumbing.

85Tool calls per prompt — the one-line delegation unit
Tooling2026-05-25

26 hands — but one of them passed 1,333 messages

Only 26 sessions opened that day. Less than 1/4 of the day before when 95 hands shared one desk. Yet those 26 hands exchanged 34,654 messages and ran 20,444 tool calls. Fewer hands, deeper hands.

1,333Messages exchanged by a single deep session
In Practice2026-05-26

A day spent building agents — the hands only touched the shell

Sessions: 74. Messages: 32,209. Tool calls: 12,878. Of those 12,878 calls, 6,382 went to one single tool — shell commands. Half of all tool use was the same gesture, over and over.

6,382Shell calls — half of all tool use that day
Failures & Cost2026-05-29

429 rate limit — the 6 minutes when the infrastructure died before the model did

Conclusion up front. An agent session tried to analyze a workspace directory tree and died 6 minutes in. The model didn't fail an answer — the request quota closed first. A cumulative 11,147 failures point at exactly that seam.

11,147Cumulative failures (7,729 sessions / 132,293 events)