humanizer

name: humanizer argument-hint: "[--strict | redo] [텍스트 또는 파일 경로] [장르: ...] [강도: ...]" description: | AI가 쓴 글의 흔적을 자연스러운 사람의 글로 바꾸는 윤문 오케스트레이터(한국어 주력). Fast(단일 호출, 디폴트)·Strict(4인 파이프라인, 정밀) 듀얼 트랙. 의미 불변·과윤문 가드·등급 자동 채점. 트리거: AI 글 자연스럽게, AI 티 제거, ChatGPT 문체, 번역투 고쳐, 사람이 쓴 것처럼 윤문, 휴머나이저, redo, 2차 윤문, --strict. group: writing model: sonnet allowed-tools: - Read - Write - Edit - Grep - Glob - Bash - AskUserQuestion - Agent

Humanizer — AI-Writing-Trace Removal Orchestrator

Finds AI-generated traces in Korean/English text and rewrites them into natural human prose. Dual track: Fast mode (default, single call) and Strict mode (4-agent pipeline).

Phase 0 — Context check and mode decision

On start, print this as the first line:

humanizer v2.0 — {fast|strict|redo} 모드 / run_id: {YYYY-MM-DD-NNN} / 언어: {ko|en|mixed}

Mode priority:

1. User says redo, or "특정 카테고리만 다시" / "이 문단만" / "2차 윤문" → redo
2. User says --strict, or "정밀 모드" / "4인 파이프라인" → strict
3. Korean input over 8,000 chars → strict auto-upgrade (one-line notice to the user)
4. English-only input → fast forced (strict is Korean-only)
5. Otherwise → fast (default)

Optional arguments (given in natural language):

• 장르: 칼럼|리포트|블로그|공적 — auto-estimated from the first 300 chars if omitted
• 강도: 보수|기본|적극 — default 기본
• 최소심각도: S1|S2|S3 (strict) or P1|P2|P3 (fast English) — default S2/P2

Phase 1 — Input capture and run_id

1. Create _workspace/{YYYY-MM-DD-NNN}/ under cwd. NNN is the day's sequence.
2. Save the input text to 01_input.txt.
3. Estimate genre from the first 300 chars (an explicit user value wins).
4. Detect language: Hangul ratio 70%+ → ko, Latin ratio 70%+ → en, else mixed.
5. On redo, reuse the most recent _workspace/ subdir when the user signals "이전 거 다시".

Fast mode (default)

Korean input

Call the humanize-monolith agent once via the Agent tool.

Call arguments:

input_path: <abs path>/_workspace/{run_id}/01_input.txt
quick_rules_path: ~/.claude/skills/humanizer/references/quick-rules.md
genre_hint: 칼럼 | 리포트 | 블로그 | 공적 | null

The monolith runs detection → rewrite → self-validation → output in one call and writes final.md + summary.md. It calls no other sub-agent.

English input (fast-only track)

The monolith is Korean-only, so the orchestrator handles English inline from this body.

1. Load references/patterns-en.md + references/patterns-common.md.
2. Decide the application bar via the content-type matrix below.
3. First-pass scan with the catalog cheat-sheet; assign severity (P1/P2/P3).
4. Fix — P1 always, P2 by context, P3 optional. Apply the 30/50% change-rate guard.
5. Run the 6-item self-check (same as Korean) — on violation, roll back that edit and retry.
6. Write final.md + summary.md. Same summary format as the monolith.

Mixed input

Do not split Korean to the monolith and English to inline handling. Process the whole text inline as the English fast track (the monolith is Korean-only and would damage English spans).

Fast response format

After writing the artifacts, return these four briefly to the user:

1. One-line status: 완료. 변경률 X% / 등급 Y / 자체검증 N/6 통과
2. The rewritten body (final.md content as a markdown block)
3. summary.md's key tables (metrics + category detection + self-check)
4. If the grade is B or lower, note "정밀 검증이 필요하면 --strict로 4인 파이프라인 실행 가능"

Default wall-clock target: ≤ 5,000 chars in 2–3 min, 8,000 chars in 5–7 min.

Strict mode (--strict or 8,000-char+ auto-upgrade)

Korean only. If --strict arrives with English input, force fast + notice: "strict 모드는 한국어 전용입니다. 영어는 fast 모드만 지원합니다."

Phase A — Detection

Call the humanize-detector agent → produces 02_detection.json. Input: 01_input.txt, genre_hint, min_severity option.

Phase B — Rewrite (up to 3 loops)

Call the humanize-rewriter agent → 03_rewrite.md + 03_rewrite_diff.json. Input: 01_input.txt, 02_detection.json, preserve_formatting option.

Phase C — Parallel verification (agent team)

Call the Agent tool twice in parallel (two tool calls in one message):

• humanize-fidelity-auditor → 04_fidelity_audit.json (13-item semantic equivalence)
• humanize-naturalness-reviewer → 05_naturalness_review.json (residual + over-polish)

Phase C verdict

On a 2nd/3rd rewrite, split versions as 03_rewrite_v2.md / v3.md. After 3 loops unresolved → hold_and_report — recommend human review.

Phase D — Final output

1. Copy 03_rewrite_vN.md or 03_rewrite.md to final.md.
2. Generate summary.md (same format as fast mode).
3. Return the four-part response to the user (same as fast).

Redo mode (/humanizer redo [instruction])

Identify the most recent _workspace/{run_id}/ (notice and exit if none).

Parse the user instruction:

Re-call humanize-rewriter + re-verify. Output is 03_rewrite_v2.md (or v3). Max round 3. Beyond that, hold_and_report for human review.

Safety net (all modes)

Invariants and guards the orchestrator applies for inline English/mixed handling and for strict verdicts. Each sub-agent also enforces its own copy inside its system prompt — that duplication is intentional, because agents run in isolated contexts and a local copy raises compliance. When changing any value below, run scripts/check-consistency to keep every copy in sync.

Supreme rule — meaning invariance

Facts, claims, numbers, dates, proper nouns, and quotations match the source 100%. Any violation triggers immediate rollback, in both the monolith and strict tracks. Genre never drifts (a column does not become an essay); register is preserved (formal stays formal — an AI tell is grammar and rhetoric, not formality itself).

Change-rate guard

• change rate = Levenshtein distance / source length.
• over 30% → warning (recorded in summary.md).
• over 50% → hard stop, roll back to the last safe version (over_polish_aborted: true).

Do-NOT list (excluded from both detection and rewrite)

• Proper nouns, product/model/org names (GPT-4, Claude 3, Gemini, etc.).
• Numbers, dates, units, percentages.
• Direct quotes inside double quotation marks.
• Legal/regulatory text.
• Math/chemistry/statistics notation.
• Industry-standard English acronyms (LLM, GPU, MCP, API, SDK, etc.).

6-item self-check (immediately after rewrite)

1. Proper nouns / numbers / dates / quotes 100% preserved.
2. Change rate ≤ 30% (over 50% = hard stop).
3. No genre drift (a column does not turn into an essay or literary piece).
4. Register preserved (formal source → formal output; do not drop to casual).
5. Zero residual S1 patterns (Korean) or P1 patterns (English).
6. No artificial expressions (no metaphor/rhetoric absent from the source added in rewrite).

On violation: roll back the edit → rewrite → re-check. Self-loop runs at most once. If still unresolved, output anyway but note "자가검증 미통과 N건" in summary.md.

Grade (auto-scored A–D)

• A: S1/P1 residual 0, S2/P2 residual ≤ 2, change rate 10–25%, self-check 6/6.
• B: S1/P1 residual 0, S2/P2 residual ≤ 4, self-check 5+/6.
• C: S1/P1 residual 1–2 or self-check ≤ 4/6 → recommend strict mode.
• D: S1/P1 residual 3+ or change rate over 50% → recommend stopping.

Operating rules

• No auto-loading. Do not auto-parse other files (project CLAUDE.md, etc.) to infer options.
• Back up prior artifacts. If final.md / summary.md already exist, back them up with a _prev suffix before writing new ones.

Content-type matrix

Genre changes the application bar. Decide in this order:

1. User specified → follow it.
2. Infer from file extension/path (e.g. README.md → technical doc, under blog/ → blog).
3. Infer from content (code-block ratio, tone, format).
4. If unclear, confirm with AskUserQuestion.

Pattern catalog reference

Read by language and mode:

patterns-ko.md has a K↔A-J mapping table at the top for cross-referencing IDs between the fast and strict tracks.

글에 숨결 불어넣기 (블로그/SNS 전용)

적용 대상: 블로그/에세이, SNS/캐주얼만. 기술 문서, 학술, 코드 주석에는 적용하지 않는다. This section stays Korean on purpose: it is domain knowledge for rewriting Korean prose and its examples are Korean output samples (see LICENSE-THIRD-PARTY for provenance).

AI 패턴 제거는 절반. 깨끗하지만 무미건조한 글도 AI처럼 보인다.

영혼 없는 글의 징후

• 모든 문장이 비슷한 길이와 구조
• 의견 없이 사실만 나열
• 불확실함이나 복잡한 감정에 대한 인정 없음
• 적절한 곳에서도 1인칭 회피
• 유머, 날카로움, 개성 부재
• 보도자료나 백과사전처럼 읽힘

숨결을 넣는 법

의견을 가져라. 사실을 보고하는 데 그치지 말고 반응하라. "솔직히 이건 좀 애매하다"가 장단점을 중립적으로 나열하는 것보다 낫다.

리듬을 바꿔라. 짧은 문장. 그리고 좀 더 천천히 가는 긴 문장. 섞어 써라.

복잡함을 인정하라. 사람은 복잡한 감정을 가진다. "인상적인데 동시에 좀 불편하다"가 "인상적이다"보다 사람답다.

'나'를 쓸 때는 써라. 1인칭이 비전문적인 게 아니다. "계속 생각나는 건..."이나 "내가 걸리는 부분은..."은 실제로 생각하는 사람의 표현이다.

약간의 지저분함을 허용하라. 완벽한 구조는 알고리즘 냄새가 난다. 곁가지, 여담, 반쯤 정리된 생각은 사람의 것이다.

감정을 구체적으로. "우려된다"가 아니라 "새벽 3시에 아무도 안 보는데 에이전트가 혼자 돌아가는 거 생각하면 좀 소름 돋는다."

References

• Korean fast rulebook: references/quick-rules.md (S1·S2 core + 6-item self-check + grade criteria)
• Korean strict SSOT: references/taxonomy-ko.md (10 categories × 40+ patterns)
• Korean strict prescriptions: references/playbook-ko.md (per-category replacement recipes)
• Korean fast catalog: references/patterns-ko.md (K1–K19 + mapping table)
• English fast catalog: references/patterns-en.md (E1–E19)
• Common patterns: references/patterns-common.md (C1–C6)

Sub-agents (used by both strict and fast)

• humanize-monolith — Fast Korean single call
• humanize-detector — Strict Phase A
• humanize-rewriter — Strict Phase B (also used by redo)
• humanize-fidelity-auditor — Strict Phase C-1
• humanize-naturalness-reviewer — Strict Phase C-2

These live in ~/.config/dotrc/agents/claude/agents/ (= ~/.claude/agents/).

Consistency check

Safety-net values (4-agent count, 30/50% guard, grade A–D thresholds, 6-item self-check, max 3 rewrite loops) are duplicated across this file, quick-rules.md, and the sub-agent definitions by design. Run scripts/check-consistency after changing any of them to catch drift.

Acknowledgements

Strict 트랙(5인 → 4인 파이프라인)과 fast 트랙(monolith)의 설계 및 quick-rules, taxonomy-ko, playbook-ko, sub-agent 정의는 epoko77-ai/im-not-ai v1.5의 humanize-korean 스킬에서 가져왔다 (MIT License). 자세한 출처와 변경 사항은 LICENSE-THIRD-PARTY에 기록.