Humanize Korean Skill (AI-tell detection and rewriting)

# Humanize Korean Skill (AI-tell detection and rewriting)

AI Summary

Purpose:

Track the humanize-korean skill (installed at

~/.claude/skills/humanize-korean/): a Korean de-AI-ification pipeline that detects and rewrites "AI tells" while keeping content byte-identical. This page records architecture facts, version history, and governance decisions that live outside the skill's own files.

Key points:

v1.5 architecture: fast mode (default) = single humanize-monolith agent

call, 3-tool-call cap, targets 2-3 min for <=5,000 chars; strict mode = 5-agent pipeline (detector, rewriter, fidelity auditor, naturalness reviewer + taxonomist off-line). Meaning invariance is the top invariant; change rate >30% warns, >50% aborts.

Taxonomy SSOT references/ai-tell-taxonomy.md is at v2.0.1 (2026-08-07,

638 lines): E-4 reinforced with the E-4/H-1 co-fire axis note, genre-guard section added (자소서: no content augmentation hard gate, A-16 first-person exempt, E-7 exempt), rejection-history section R-1..R-5 added. Pattern IDs and severities unchanged; v2.1 reserved for the A-17 NMT round.

Promotion discipline (proven useful 2026-08-07): external tips enter via

taxonomist adjudication (accept/reinforce/reject with verbatim-span evidence gate), then quick-rules-integrator lands rulebook changes under hard guards (180-line cap, monolith definition frozen, no push without user approval). A viral YouTube 자소서 tip yielded 0 new patterns but exposed and fixed 4 pattern-ID desyncs between quick-rules and the SSOT.

Key doctrine from that audit: "AI-authorship tells" and "bad writing" are

different axes — signals humans overuse too (effort/growth clichés, result-only narration) have no discriminative power and must not become patterns; content gaps in cover letters are detect-and-report only, never filled by the rewriter.

Relevant when:

Upgrading or debugging the humanize-korean skill; adjudicating new AI-tell

pattern candidates; answering why a candidate was rejected; planning the v1.6.2 / v2.1 rounds.

Do not read full document unless:

You need the exact audit trail or follow-up details.

Linked documents:

ai/worklog/2026/2026-W32.md (2026-08-07 audit entry)
~/.claude/skills/humanize-korean/references/ai-tell-taxonomy.md (SSOT)
~/.claude/skills/humanize-korean/CHANGELOG.md

Open Questions

Whether the 6 S1-vs-S2 severity uplifts in quick-rules are intentional

fast-path design or drift (no written rationale found).

Where the skill's canonical git remote lives; this machine's install is

not a repo, so cross-machine version sync path is Needs confirmation.

Details

2026-08-07 — YouTube 자소서-tip audit (v2.0.1)

Source: "GPT 쓴 티나는 자소서 특징 5가지" (youtube.com/watch?v=J9XC2JuCmD8), anecdotal, genre-specific. User instruction: keep what survives, discard the rest. Pipeline: korean-ai-tell-taxonomist (adjudication + SSOT edit) → quick-rules-integrator (rulebook landing + regression guards).

Verdicts:

#	Claim	Verdict
1	Fragmented sentences, no connectives	Reinforce E-4; not a

contradiction of H-1/A-14 but the same deficit's two faces (missing connective ENDINGS, patched with sentence-initial connective ADVERBS); they co-fire. Remedy order: delete adverbs first, then merge clauses. | | 2 | Effort/growth clichés, no substance | Reject — content defect, human cliché, no discriminative power. | | 3 | Comma overuse | Reject — C-11/C-12/E-5/E-6 already strictly stronger. | | 4 | Mechanical listing | Reject — C-1/C-2/C-9/A-13 already cover it. | | 5 | Result-only narration, process omitted | Reject as pattern; intent inverted into the 자소서 hard gate (content gaps: detect-and-report only). |

The tip's bundled single-prompt "review prompt" was discarded as harmful for the genre: it instructs rewriting against criteria 2/5, which invites fabricating applicant experience.

Files changed in the skill: references/ai-tell-taxonomy.md (v2.0.1, 638 lines), references/quick-rules.md (140/180 lines, 4 ID desyncs fixed, C-1 [S1]/E-4/C-12 added, 자소서 guard), references/rewriting-playbook.md (§4 자소서 row), SKILL.md (genre enum), CHANGELOG.md (new). PR draft under the skill's _workspace/v2.0.1-2026-08-07/07_pr/ — not pushed.