humanizer-kill-gpt

name

humanizer-kill-gpt

description

Runs a cross-language cleanup for writing tells shared by large frontier LLMs, not only GPT: English source constructs that survive Korean translation and Korean-surface translationese. Triggers on de-GPT, GPT 티, AI 티, 번역체 GPT, 자폐 말투 (colloquial shorthand for rigid or over-complete model prose, not a claim about autism), or English text bound for Korean; broad Korean punctuation and spacing cleanup stays out of scope.

Humanizer: Kill GPT Tells

Remove recurring AI-writing fingerprints from English and Korean text. Despite the skill name, these patterns are not unique to GPT: large frontier language models commonly converge on the same over-complete, over-explained, overly balanced, and translation-shaped prose. The catalog keeps “GPT tell” as a familiar shorthand.

Some users call this style 자폐 말투. Treat that phrase only as colloquial request vocabulary for rigid or socially miscalibrated model prose. MUST NOT equate the writing patterns with autism or use the phrase to characterize autistic people.

Most translated Korean tells start upstream: an English model-writing construct survives literal translation. This skill therefore uses two coupled passes.

Text situation	Run
English draft	Pass A only
Korean draft written natively	Pass B only
Korean translated from English, or English that will be translated to Korean	Pass A first, then Pass B

This skill targets cross-language patterns shared by large frontier models. A broader Korean-language cleanup workflow may cover additional punctuation, spacing, and lexical-distribution patterns that are outside this skill's scope.

When the user asks for generation-time speaking guidance instead of post-editing, read references/generative-speaking-directive.md. Do not invert the catalog into a long prohibition list. Correct the three root behaviors instead: exhaustive coverage, over-helping, and formal-writing defaults.

Procedure

Input contract

Resolve INPUT_TEXT before [S1].

• SKILL_DIR MUST be the absolute directory containing this SKILL.md. If the host exposes only an installed alias, run realpath "{installed-skill-path}" and use the returned directory.
• Preserve the source line breaks and number them from 1 for findings output.
• Initialize A_HITS=[] and B_HITS=[] before route-specific scans.

Input condition	Exact action
Text is present in the current conversation	Use that text verbatim as INPUT_TEXT; no external tool call is required
User supplies a file path	Run sed -n '1,240p' "{path}"; continue with sed -n '241,480p' and later ranges until EOF
No text or file is available	MUST ask one short question requesting the text, then wait

Step 1 [S1]: Select the execution route

Tool: Current-context language classification; no external command or model.

Action: MUST classify INPUT_TEXT with the decision table below. Use in-context language analysis; MUST NOT call an external rewriting model.

Condition	Set ROUTE	Required path
English text that stays English	A	[S2] → [S3] → [S5]–[S9]
Korean text written natively	B	[S2] → [S4]–[S9]
Korean translated from English	A+B	[S2]–[S9]
English text intended for Korean readers or later translation	A+B	[S2]–[S9]; [S4] performs a prospective Korean-tell check
User requests generation-time persona or speaking guidance instead of post-editing	DIRECTIVE	[S2] → [S9]

VERIFY [S1]: A working line exists in the form Route: {A|B|A+B|DIRECTIVE} and exactly one route is selected.

Step 2 [S2]: Load the route references

Tool: The host's file-read tool; fallback to the shell command sed.

Action: MUST read the files selected below from SKILL_DIR, the directory containing this SKILL.md.

ROUTE from [S1]	Exact file-read path
A	{SKILL_DIR}/guides/english-source-constructs.md
B	{SKILL_DIR}/guides/natural-korean-patterns.md
A+B	Read both guide paths above
DIRECTIVE	{SKILL_DIR}/references/generative-speaking-directive.md

IF BLOCKED: If the host has no file-read tool, MUST run sed -n '1,240p' "{exact-path}" and continue with later 240-line ranges until EOF. If a guide still cannot be read, use the inline catalog below and disclose the missing guide in the final response.

VERIFY [S2]: Every file required by the selected route was read through EOF, or its missing-path fallback was recorded.

Step 3 [S3]: Scan English source constructs

Tool: Current-context sentence comparison; no external rewriting model or web service.

Action: For ROUTE=A or ROUTE=A+B, MUST compare INPUT_TEXT against all 14 Pass A body rows exactly once, in table order. Use in-context text analysis; MUST NOT rewrite during this step. Store each match as A_HITS with fields line, source span, pattern, reason, and Korean consequence.

VERIFY [S3]: The working record states A_ROWS_CHECKED=14; A_HITS contains one record per Pass A match, including zero records when no match exists.

Step 4 [S4]: Scan Korean-surface tells

Tool: Current-context sentence comparison; no external rewriting model or web service.

Action: For ROUTE=B or Korean ROUTE=A+B, MUST compare INPUT_TEXT against B1–B7 exactly once, section order first and row order second. For English ROUTE=A+B, MUST instead test whether a literal Korean translation of each sentence would regenerate any B1–B7 tell. Store matches as B_HITS with fields line, pattern, and reason.

VERIFY [S4]: B_HITS contains one record per applicable B1–B7 match, including zero records when no match exists; every applicable section was checked.

Step 5 [S5]: Build the findings table

Tool: Current-context table synthesis; no external tool.

Action: MUST assign every hit DEDUP_KEY={line}|{source span}, merge A_HITS and B_HITS by that key, and produce the exact structure below. source span means the smallest contiguous substring that triggered one or more categories; MUST NOT widen it to the whole line. Overlapping or nested hits share one key, while non-overlapping hits on the same line keep separate keys. Combine every category and reason attached to the same key in one row. Each replacement MUST preserve facts, certainty, intensity, register, and intentional code-switching.

| Original line | Pass/category | Why it reads as model-written | Proposed replacement |

For translation-bound text, the same row MUST name the English source construct and the Korean tell it would produce. A proposed replacement MUST be checked against the full catalog before inclusion.

VERIFY [S5]: The table row count equals the count of distinct DEDUP_KEY values; no key appears twice, and every row has a concrete replacement that introduces no cataloged tell.

Step 6 [S6]: Apply the requested edit mode

Tool: Current-context text rewriting; no file-write tool unless the user explicitly requested an in-place file edit.

Action: MUST select one mode from the decision table and produce EDITED_TEXT.

Condition	Required action
User explicitly requests no edits, review only, findings only, or diff only	Set EDITED_TEXT=INPUT_TEXT; preserve the findings table as the proposed diff
Findings exist and review-only mode was not requested	Apply every proposed replacement to a full rewritten text
No findings exist	Set EDITED_TEXT=INPUT_TEXT without cosmetic rewriting

VERIFY [S6]: EDITED_TEXT exists, preserves the source meaning, and reflects exactly one mode from the table.

Step 7 [S7]: Run the deterministic residual gate

Tool: Execute the bundled scripts/residual_check.py with python3 "{SKILL_DIR}/scripts/residual_check.py". Replace {SKILL_DIR} with the absolute path resolved in the input contract. Reading the script is unnecessary but MUST NOT replace execution.

Action: MUST map the findings-table categories to the deterministic IDs below, deduplicate the IDs, and pipe the complete EDITED_TEXT from [S6] to the script. In review-only mode, pipe the proposed replacement text instead of unchanged source lines. The residual-check command MUST be the first and only shell invocation in this step. Keep the JSON result in the current response context; MUST NOT write EDITED_TEXT, script output, or intermediate findings to any file, including a temporary file.

Findings category	Deterministic ID
Negative parallelism	negative-parallelism
Rule of three	rule-of-three
Copula avoidance	copula-avoidance
Fit assertion	fit-assertion
Significance inflation	significance-inflation
-ing tail clause	ing-tail
Corporate softener	corporate-softener
Signposting	signposting

Set CATEGORY_FLAGS to one --category {deterministic-id} flag per unique mapped ID. If the findings contain none of these categories, set CATEGORY_FLAGS=--category none.

python3 "{SKILL_DIR}/scripts/residual_check.py" {CATEGORY_FLAGS} <<'EDITED_TEXT'
{complete EDITED_TEXT or proposed replacement text}
EDITED_TEXT

The script emits one JSON object and uses exit codes 0=pass, 1=residual found, 2=invalid invocation or input.

Script result	Required action
Exit 0, JSON status=pass	Set deterministic_remaining=0; continue to [S8]
Exit 1, JSON findings present	Rewrite only the reported spans, then repeat [S7] once
Second exit 1	Set deterministic_unresolved to the reported findings; go directly to [S9] and disclose them
Exit 2	Fix the invocation or input named in JSON, then repeat [S7] once
Second exit 2	Go directly to [S9] and disclose that deterministic verification failed

IF BLOCKED: If the direct python3 command exits 127, run the same command once with python. If that also exits 127, go directly to [S9] and disclose that the deterministic gate was unavailable.

VERIFY [S7]: Before [S8], the residual-check command exits 0; JSON has status=pass and finding_count=0; checked_categories exactly matches the unique mapped IDs, or is empty only when --category none was used; S7 used no ls, stat, which, or artifact-write call.

Step 8 [S8]: Re-scan the full catalog

Tool: Current-context sentence comparison using the same catalogs loaded in [S2].

Action: MUST compare every proposed replacement first with the category that triggered its row, then repeat the applicable [S3] and [S4] comparison order against EDITED_TEXT. In review-only mode, scan the proposed replacement text. A synonym that retains the same structure MUST count as same_category_remaining.

Re-scan result	Required action
same_category_remaining=0 and no other avoidable tell remains	Continue to [S9]
A replacement retains its source category or introduces another tell	Rewrite only that replacement, then repeat [S7] → [S8] once
The repeated [S7] or [S8] still fails	Set unresolved to the remaining items and continue to [S9] without claiming a clean result

VERIFY [S8]: The re-scan record lists same_category_remaining, remaining, introduced, and unresolved; a clean result requires the first three counts to equal 0.

Step 9 [S9]: Deliver the result

Tool: The host's final-response renderer; MUST NOT write an artifact unless the user requested one.

Action: MUST emit exactly one output shape from the decision table.

Condition	Output shape
ROUTE=DIRECTIVE	Generation-time directive from the loaded reference, followed by one short rationale
No findings	Route: {ROUTE} followed by GPT/frontier-model tells: none
[S7] verifier unavailable or errored twice	Route line → findings table → proposed or edited text → verifier error; MUST NOT claim a clean re-scan
Review-only mode	Route line → findings table → proposed diff or replacement text → deterministic and full-catalog forecast
Applied mode	Route line → findings table → complete EDITED_TEXT → Re-scan: deterministic_remaining=0, same_category_remaining=0, remaining={N}, introduced=0, unresolved={N}

VERIFY [S9]: The response contains the selected route and required artifacts. Post-editing routes include the actual deterministic plus full-catalog status without overstating success; DIRECTIVE omits scan status by design.

MUST NOT replace one model-writing tell with another. MUST preserve all existing functionality and the user's intended voice.

---

Pass A: English source constructs

These English-side habits occur across large frontier LLMs and produce recognizable Korean translationese. Removing them in English is the cleanest fix. See guides/english-source-constructs.md for detailed before-and-after examples.

Negative parallelism is the highest-priority pattern to rewrite.

English GPT construct	Why it is a tell	Korean tell it becomes	Fix in English
Negative parallelism: "not just X, it's Y" or a "no guessing" tail	Overused contrast frame	~가 아니라 ..., ~하기보다, or a ~없이 tail	State the positive clause directly and give the reason
Copula avoidance: "serves as / stands as / acts as / functions as a X"	Inflated alternative to plain "is"	~역할을 합니다, ~로 자리합니다	Use "is" or "is the"
Fit assertion: "X is a natural fit for Y"	Asserts suitability through another copula-avoidance frame	~에 자연스럽게 맞습니다, ~가 잘 맞습니다	Use "X fits Y" or "use X for Y"
Rule of three: "fast, simple, and reliable"	Forced completeness through a triad	GPT-style three-noun list	Keep only the one or two items that matter
Colon apposition: "the official community: a moderated place where..."	Restates the noun with artificial depth	~하는 공간으로 padding	Split it into plain sentences
-ing tail clauses: "..., ensuring/highlighting/reflecting X"	Adds bolt-on significance	~을 보장하며, ~을 반영하며	Make it a separate sentence or delete it
Hedging stack: "it could potentially be argued that it might..."	Stacks unnecessary qualification	~일 수도 있다고 볼 수도 있습니다	Hedge once, or state it directly
Corporate softener: "whatever you need", "feel free to", "I'd be happy to"	Servile padding	부탁드립니다, 혹시 괜찮으시다면 padding	Delete it; keep at most one polite marker
Significance inflation: "marks a pivotal moment", "underscores its importance"	Inflates an ordinary fact	중대한 전환점, ~의 중요성을 보여주는	State the fact or cut the sentence
Em-dash overuse	Creates synthetic sales-copy rhythm	Repeated dashes or awkward translated pauses	Use commas, periods, or parentheses
Signposting: "Let's dive in", "Here's what you need to know"	Tutorial-script warm-up	자, 그럼 ~, 여기서 알아둘 점은	Start with the substance
Exhaustive if-else branching	Covers every branch instead of committing to the likely one	~했다면 ..., 안 했다면 ...	Choose one branch or ask one short question
Contraction avoidance in casual or spoken English	Sounds robotic in small talk and voice output	English-only tell	Use don't, didn't, and can't
Small-talk-to-coaching conversion	Turns a greeting into unsolicited advice	Advice appended to a simple greeting	Reply briefly, match the register, and return the question

Pass A self-check: Translate each rewritten English line literally. If a Pass B tell comes back, rewrite the English again.

---

Pass B: Korean-surface GPT tells

Run this pass on Korean text, whether native or translated. The tables identify what to remove; guides/natural-korean-patterns.md provides natural alternatives and rhythm.

B1. Non-native Korean

These machine-translated structures sound immediately unnatural to native speakers.

GPT expression	Problem	Natural direction
<English word> 계열 <English word>	Mixes English terms through the vague noun 계열	Translate the relationship or use ~ 종류의
A 없음은 B를 proxy로 봤습	Non-native nominalization and case marking	A가 없어서 B를 대신 봤음
흐름 안에서	Habitual insertion in an unnatural position	Delete it, or use ~ 과정에서 / ~ 중에
...은 받지 않겠습니다	Needlessly hard and imperious	Use ~는 어렵습니다 or ~는 수용하기 힘듭니다 as context requires
다른 분들 평가서에 보면...	Non-native -서에 보면 construction	다른 분들의 평가를 보면

B2. Grammatically valid but GPT-shaped Korean

GPT expression	Problem	Natural direction
Repeated topic-marker forms such as ...에서는, ...보다는, ...으로는	Translationese topic marking	Rebuild the sentence and remove unnecessary 는 forms
...으로는 ...지만, ...지는 않았습니다	Double reservation	Assert once or hedge once
-로 봤습니다	Overused judgment phrase	~라고 판단했습니다, ~로 생각했습니다, or a direct assertion
~가 자연스럽게/잘/딱 맞습니다 as a suitability claim	Literal "is a fit" construction	~가 제격입니다, ~면 됩니다, ~로 가면 됩니다
...정도로 넓혀도 N건 수준이라	Stacks 정도 and 수준	State the number directly
넓히다, 좁히다 for abstract scope	Overused metaphor	늘리다/줄이다, 확대하다/축소하다
열다, 닫다, 굴리다 for discussions or businesses	Overused physical metaphor	시작하다, 마무리하다, 정리하다, 다루다
두께, 두텁다 for abstract data, relationships, or experience	Literal "depth/thickness" metaphor	Use 양, 깊이, or explain what accumulated
척추, 등뼈 for a business or system	Literal body metaphor	근간, 뼈대, 토대, 기반
-지는 않았습니다	Unnecessary reservation	End with a direct assertion
...쪽이었고	Padding through 쪽	Use ~였고
~한 거 맞음 as self-confirmation	The speaker re-confirms their own statement	Delete the confirmation tail
니 말이 맞음, 말씀이 맞습니다, 말씀하신 대로 as an opener	Reflexive agreement and flattery	Start with the evidence or the answer
하나로 묶다	Overused abstraction	합치다, 한데 모으다
...로 봅니다, ...로 보지 않고	Repetitive translated judgment frame	Use natural spoken Korean or a direct assertion
그 위에서, ... 위에서 with an abstract object	Physical positioning attached to an abstraction	그걸 바탕으로, 그 다음에
다투다 in neutral competition contexts	Overused dramatic verb	겨루다, 경쟁하다, 부딪히다
그 지점은	Habitual abstraction	그 부분은
-하고자 합니다	Stiff intention form	제안드립니다, 말씀드립니다
~기가 쉽지 않은 상황입니다, ~상황입니다	Adds a padded situation noun	State the difficulty with a verb
~점도 충분히 이해하고 있습니다	Scripted empathy padding	Use brief, context-specific empathy only when needed
가능한 범위 안에서	Stock phrase	가능한 선에서, 가능한 선까지
박하게 with estimates or offers	Often sounds forced	Use 낮게 or explain the decision
Six or more nouns in a list	Exhaustive cataloging	Keep three or four representative items
~게 아니라 ~니다	Negative parallelism preserved in Korean	Split it or state the positive claim directly
토대/기반 위에 세우다, 그 위에서	Literal construction metaphor	~를 바탕으로, ~에서 출발하다
레이어 for abstract capabilities	Literal "layer" metaphor	Explain the capability, use 기능, or delete it
해자 in ordinary business prose	Literal "moat" jargon	Explain the barrier directly
묶이다, ~에 묶입니다 as "bound to"	Literal English mapping	~를 벗어나지 못하다, ~ 안에서만 돌아가다
A full root-cause analysis in response to a one-line status question	Over-explanation is itself an AI tell	Answer 지금은 정상, 재현 안 됨; explain only when asked

B3. Corporate softeners inserted without need

Delete ceremonial padding that was absent from the source:

• 부탁드립니다, 양해 부탁드립니다
• 미리 감사합니다, 감사합니다 (미리)
• 말씀 드리고 싶었습니다
• 혹시 괜찮으시다면

Keep one polite marker only when the context genuinely requires it.

B4. Hedging adverbs

Check 아마, 혹시, 어쩌면, 다소, and 충분히. Remove them when the source makes a direct claim. Preserve one only when the uncertainty is real.

B5. AI-slop and robotic markers

Marker	Problem	Alternative
·	Machine-generated visual texture	Comma, line break, or 와/과
Honorifics applied to objects	Incorrect elevation	Use the ordinary verb form
-십시오	Mechanical register	-세요
Sentence-final -니까 in a question	Awkward ending	-나요? when appropriate
당신	Translated second person	Use a name, handle, or omit the pronoun
-합시다	Mechanical suggestion	-하면 좋겠습니다, -해주세요, 어떻게 생각하세요?
-라고 봅니다	Translationese judgment	-라고 생각합니다 or a direct assertion
-쪽, -쪽으로, -쪽입니다	Padding	Delete it or use ~로

B6. Artificial narrative structure

These patterns are structurally theatrical even when each word is grammatical.

• Narrative verbs for plain facts: Replace unnecessary 들어갑니다, 나갑니다, and 넘어갑니다 with 시작합니다, 추가됩니다, or a direct state. Example: 신제품은 7월입니다, 그다음은 이 제품군 밖입니다.
• Physical motion assigned to abstract actors: Replace 저희 도구는 작업 안에서 움직입니다 with the actual meaning, such as 저희 도구는 작업 중에 실시간으로 함께 작동합니다.
• Dramatic subject reveal: Put the subject before the judgment. Replace 그중 제일 잘한 건 인정합니다. 제품 X는... with 그중 제품 X는 솔직히 잘 만들었습니다.
• Dropped object particles: Restore 을/를 when headline-style omission makes the sentence sound generated: 보이스 라인업을 준비 중입니다.

B7. Defensive echoing

Do not delay a new answer by repeating the user's premise or defending the previous answer.

• Echoing the user: Delete openers such as 말씀하신 대로 ~라면 and answer directly.
• Re-explaining the previous answer: Remove the defense and address the new question. If context acknowledgment is necessary, keep it to one short clause.

Preservation rules

Do not damage natural text through over-correction:

• Preserve intentional loanwords and code-switching such as 프로세스, 톤, and 내로남불.
• Preserve genuine rhetorical questions such as 내가 왜 이해해줘야 함?.
• Preserve the source's certainty and intensity. Do not add corporate softeners.
• Preserve a consistent nonstandard spacing pattern or voice when it is intentional.
• In English, preserve an intentional single em dash, real terms of art, and the author's own asserted strength.

The mandatory output contract is defined in [S5] and [S9]. MUST NOT substitute another format.