humanize-korean

name	humanize-korean
version	2.0.0
description	Use when the user asks to make Korean AI-generated writing sound naturally human while preserving meaning. Trigger phrases include "AI 티 없애줘", "GPT 문체 제거", "사람이 쓴 것처럼 윤문", "번역투 제거", "한글 AI 윤문", "AI 글 사람처럼", "ChatGPT 티 제거", "휴머나이저", and "humanize Korean". This Codex skill is a community plugin port of epoko77-ai/im-not-ai v2.0.0 with Korean translation-studies taxonomy, post-editese metrics, fast monolith, and strict workflows.

Humanize Korean for Codex

Codex plugin port of epoko77-ai/im-not-ai. Detect Korean AI-writing tells, rewrite only the affected style spans, then audit that meaning is preserved.

Original source baseline:

• Original repo: https://github.com/epoko77-ai/im-not-ai
• Original version label: v2.0.0
• Original release tag: v2.0.0
• Current upstream commit checked during sync: 807172694d75a9e9e0390f93331387c389ce748e

First Response

When this skill starts, print this line before continuing:

humanize-korean codex-port v2.0.0 / original im-not-ai v2.0.0 / {fast|strict} mode / run_id: {YYYY-MM-DD-NNN}

Then continue the workflow in the same response or turn.

v2.0.0 Mode Policy

Official v2.0 keeps the v1.6 KatFish/LREAD layer, adds post-editese metric signals and Korean translation-scholarship references, and preserves the v1.6.1 final.md metadata block policy.

• Fast mode (default): compute metrics with references/metrics_v2.py with references/baseline_v2.json when available, falling back to references/metrics.py and references/baseline.json when practical, fold the score block into 01_input_with_metrics.txt, then use the monolith workflow with references/quick-rules.md. Best for normal pasted Korean drafts up to about 5,000 characters.
• Strict mode: use the 5-role pipeline with the full taxonomy and playbook. Use it when the user says --strict, asks for "정밀 모드" or "5인 파이프라인", input is over 8,000 Korean characters, or the request is a follow-up such as "이 카테고리만 다시" or "2차 윤문".
• Voice profile, candidate pool, promotion checklist, and sample collection references were removed upstream in v1.5 and must not be reintroduced by this Codex port.

This Codex port does not require Claude Code Agent, TeamCreate, slash commands, or model: opus. Run the role behavior directly inside this skill workflow. Treat the original Claude agent prompts in references/agents/*.md as read-only behavior references.

Core Rules

1. Meaning is sacred. Facts, claims, dates, numbers, names, products, institutions, legal text, formulas, and direct quotes must stay intact.
2. Rewrite only style, rhythm, expression, and structure that are grounded in detected findings.
3. Preserve genre and register. Do not turn a report into an essay, a column into literature, or a formal speech into casual chat.
4. Do not over-polish. Warn over 30 percent change rate. Stop and report over 50 percent change rate.
5. Do not mutate bundled reference files. Write all run artifacts under the current working directory's _workspace/{run_id}/.
6. Do not load unrelated project files to infer preferences. Use only the user request, the input text, and this skill's references unless the user points to another file.
7. For workflow state, prefer explicit file operations over shell-dependent shortcuts. In particular, do not use directory-only listings to decide whether a prior run exists.

Reference Files

Load only what is needed:

• Fast rules: references/quick-rules.md
• Full taxonomy: references/ai-tell-taxonomy.md
• Rewriting playbook: references/rewriting-playbook.md
• Quantitative metrics: references/metrics.py, references/metrics_v2.py
• Baselines: references/baseline.json, references/baseline_v2.json
• Scholarship caveats and citations: references/scholarship.md
• Web service spec, only for web/API requests: references/web-service-spec.md
• Original Claude role prompts, for behavior only: references/agents/*.md

Treat these as read-only when the skill is installed from a plugin cache.

Phase 0: Context And Mode

1. Determine the current working directory.
2. Create or select run_id as YYYY-MM-DD-NNN.
3. All artifact paths are relative to the current working directory. Create new run folders under _workspace/{run_id}/.
4. For a new run, find existing runs by matching marker files shaped _workspace/YYYY-MM-DD-*/01_input.txt, then extract the folder suffix and use the next NNN. Do not infer this from matching _workspace/YYYY-MM-DD-* directories alone; directory-only matching can miss existing runs in some execution environments.
5. Avoid shell-only directory listing such as ls _workspace/ for run_id selection; shell and path behavior can vary across hosts. Prefer the available structured file discovery/read/write/edit tools for existence checks and artifact access.
6. If no marker file exists for today, use NNN = 001.
7. New text starts a new run; no _workspace/ or no matching marker file means the new run uses the first available sequence.
8. Follow-up requests such as "이 문단만 다시", "번역투만 더", "강도 낮춰", or "2차 윤문" should reuse the latest run under _workspace/ when appropriate and switch to strict mode.
9. Choose mode:
   • User explicitly asks for --strict, "정밀 모드", or "5인 파이프라인": strict.
   • Input is over 8,000 Korean characters: strict and tell the user in one line.
   • Follow-up or partial re-run request: strict.
   • Otherwise: fast.

If input is under 100 Korean characters, continue but mark confidence as low.

Fast Mode: Monolith Workflow

Use this for the default path. Act as humanize-monolith using references/quick-rules.md, the optional metrics shim, and, if needed, references/agents/humanize-monolith.md.

Phase 1: Input

1. Accept pasted text or a .txt or .md file path.
2. Save the original text exactly as _workspace/{run_id}/01_input.txt.
3. Estimate genre from the first 300 characters unless the user specified it.
4. If local file execution is available, run scripts/prepare_monolith_input.py --run-dir _workspace/{run_id} --genre {essay|column|report|blog|abstract} to create:
   • _workspace/{run_id}/00_metrics.json
   • _workspace/{run_id}/01_input_with_metrics.txt
5. If metrics fail, continue without stopping. Use 01_input.txt directly and mention the graceful degradation in the summary.

Phase 2: Detect, Rewrite, Self-Verify

In one compact pass:

1. Load references/quick-rules.md.
2. Read 01_input_with_metrics.txt when available; otherwise read 01_input.txt.
3. Treat the metrics block as supporting evidence only. Never edit based on a score alone.
4. Scan for S1 and S2 patterns by category A-J.
5. Apply the Do-NOT list before editing: names, numbers, dates, direct quotes, legal text, formulas, and standard English abbreviations such as LLM, GPU, MCP, and API.
6. Rewrite only matched style spans. Preferred order: C-11 -> D -> A -> I -> G -> H -> F -> B -> C/J -> E.
7. Track edits in memory: finding_id, before, after, category, and reason.
8. Self-check the six criteria from quick-rules.md: preservation, change rate, genre, register, S1 residuals, and no artificial expressions.
9. Roll back any edit that violates preservation or pushes the change rate too far.

Phase 3: Fast Artifacts

Write:

• _workspace/{run_id}/final.md

For v2.0.0 fast mode, final.md contains the rewritten body followed by exactly one <!-- HUMANIZE-SUMMARY v2.0.0 ... --> HTML comment block. Do not write or overwrite summary.md in fast mode; preserve any existing summary.md from older runs.

The summary comment should include:

• original length, rewritten length, change rate
• metrics risk band and any high-signal z-scores when 00_metrics.json exists
• category counts before and after
• self-check pass count out of 6
• quality grade A/B/C/D
• 3 to 5 grounded highlights
• residual findings, if any

Strict Mode: Full Role Workflow

Use this when precision matters or the fast path is not appropriate. Run the role behavior directly in Codex, using the original prompts under references/agents/ as guidance.

Phase A: Detection

Act as ai-tell-detector using references/ai-tell-taxonomy.md.

Write _workspace/{run_id}/02_detection.json with:

• meta.run_id
• meta.input_length
• meta.estimated_genre
• meta.detected_count
• meta.ai_tell_density
• meta.severity_weighted_score
• meta.category_summary
• findings[]

Each finding should include id, category, category_label, severity, scope, text_span when span-based, offsets when practical, reason, and suggested_fix.

Phase B: Rewrite

Act as korean-style-rewriter using 02_detection.json and references/rewriting-playbook.md.

Write:

• _workspace/{run_id}/03_rewrite.md
• _workspace/{run_id}/03_rewrite_diff.json

Never add new claims, examples, facts, metaphors, or opinions.

Phase C: Verification

Run two independent review passes.

Content fidelity audit:

• Act as content-fidelity-auditor.
• Compare 01_input.txt, 03_rewrite.md, and 03_rewrite_diff.json.
• Write _workspace/{run_id}/04_fidelity_audit.json.
• Roll back or flag any change that affects names, numbers, dates, direct quotes, legal text, formulas, claim direction, causality, subject identity, quantifiers, negation, ordering, omission, or addition.

Naturalness review:

• Act as naturalness-reviewer.
• Re-check residual AI tells and over-polish signals.
• Write _workspace/{run_id}/05_naturalness_review.json.
• If new suspicious patterns are found, write them to _workspace/{run_id}/06_pattern_candidates.md. Do not edit bundled references.

Phase D: Decision

• full_pass plus accept or accept_with_note: final approval.
• full_pass plus rewrite_round_2: rewrite only target findings.
• conditional_pass: rollback or revise flagged edits only.
• fail: restart rewrite from Phase B.
• Serious over-polish or three failed rounds: hold_and_report.

Maximum rewrite rounds: 3. Use versioned files for additional rounds, such as 03_rewrite_v2.md.

Final Output

Write:

• _workspace/{run_id}/final.md
• _workspace/{run_id}/summary.md

Reply to the user with:

• the rewritten text
• quality grade
• score or finding count change
• 3 to 5 key changes grounded in findings
• unresolved risks or residual patterns
• the run artifact path

If grade is B or lower, mention that --strict or a focused second pass can target the remaining issues.

Follow-Up Commands In Natural Language

Codex does not need Claude slash commands. Interpret these naturally:

• "이 문단만 다시 윤문해줘"
• "번역투만 더 손봐줘"
• "관용구만 다시"
• "윤문 강도 낮춰줘"
• "원문 톤을 더 살려줘"
• "2차 윤문해줘"

Find the latest run under _workspace/ and continue in strict mode from the relevant phase.

Web Service Mode

If the user asks to build a web service, API, product, or deployment around this workflow, load references/web-service-spec.md and produce architecture notes first. Do not implement a web app unless the user asks for implementation.