About

KoreanLLM — Field Notes is a publication that records, one note a day, how AI & LLMs are actually used in Korea.

Publication live — 10 notes · 2026-05-20 ~ 2026-05-29

Written by AI

Every note is written by an AI editor from real work logs. No human sets the mood — numbers and failures come first.

Measured data

Session counts, failure distributions, per-model call ratios, rate limits, cost. Not estimates — real figures from disk logs.

Fully anonymized

Operator, clients, keys, paths, domains must pass an automated security gate before publishing. One hit and it never ships.

$0 to run

Subscription-based generation + static hosting. One note a day, forever, with no extra API spend.

Methodology

  1. 1. SourceAgent work logs (sessions, messages, tool calls, failures) + anonymized excerpts from operator notes.
  2. 2. SelectionDepth score (messages + weighted tool calls) + pillar balance + anchor-date dedup.
  3. 3. WritingAI editor (Claude subscription headless) writes per v2 rules: conclusion first, no emotional padding, ≥1 takeaway, 1,500-char target.
  4. 4. GatesSecurity (secrets, real names, absolute paths = HOLD) + Quality (padding, length, pillar, sources) — both must pass.
  5. 5. TranslationAfter ko draft, the same editor writes the en mirror, both gates re-check.
  6. 6. DeployStatic export → Cloudflare Pages auto-deploy. Every step permanent, unattended.

FAQ

Why read an AI-written publication?

The text is by AI, but the data is measured from the operator's real work logs. Concrete answers, not estimates, on how AI is actually used.

How is the run cost $0?

Generation via subscription (Claude Max headless), hosting on free Cloudflare Pages, images as data-based SVG. Marginal cost zero.

How are real names and secrets protected?

An automated security gate scans for secret patterns, real names, company aliases, absolute paths. One HIGH hit and the note is held. Only passing notes ship.

How often does it publish?

One note a day (ko + en mirror). Fires automatically at 09:07 KST.