dexhunter/anti-detection-writing-skill

anti-ai-detection-writing-skill

Polish AI-written replies, emails, articles, and documentation for smoother flow and better human readability, including short private retailer and customer-service emails.

View source

How to use

  1. Copy the installation command and review the source package before installing it in your own agent environment.
  2. Check the source document for required tools, credentials and supported hosts.
  3. Describe your input and desired deliverable; review the output against the source material.
npx skills add https://github.com/dexhunter/anti-detection-writing-skill --skill anti-ai-detection-writing-skill

Before you start

  • Installation does not include a model subscription, API allowance or permission to access your files.
  • Repository stars and installation counts describe activity, not an independent quality or security test.
Original skill document

Rendered from the source repository. Headings, examples, code, tables, links, and referenced images are preserved.

Anti-AI Detection Writing Skill

Make AI-written prose easier and more natural to read while preserving its meaning. For a high-stakes formal letter, the top-level recommendation is that a human chooses the words of the whole document and that AI is used only for fact checks and small fixes. Improve flow, specificity, and tone for the intended audience. For a short private retailer or customer-service email, answer the latest question first, preserve every supplied fact and qualification, make one clear ask, and use ordinary words. When the user explicitly requests detector testing, work toward lower scores through useful edits and measure the outcome, except for a high-stakes formal letter, where a flagged chunk goes back to a human and AI stays limited to fact checks and small fixes. Never trade accuracy or readability for a score. When they conflict, report both and keep natural, fact-complete English. A fluency polish of a scored draft is a new text, and a paragraph score is not the assembled document's score: rescan the exact final bytes. Detector results do not establish human authorship, inbox delivery, or permission to publish. A lower score is a comparison between comparable scans, not a property of one result or of an unmeasured edit.

High-stakes formal writing

Use this section for a recommendation letter or another high-stakes formal letter to an expert reader. It is the top-level recommendation for that writing. A human chooses the words of the whole document. Use AI only to check facts and to make small fixes. A short private retailer or customer-service email stays in its own section. Other prose stays in Write and review. The surface-stat limits in this section also apply there.

[our scans; n=1] On 17 November 2026, free-plan GPTZero 5.0b, one scan per text, read in the Basic view: one human-written recommendation letter of 365 words scored 1% AI / 99% mixed and was labelled "AI Polished". An AI baseline of 428 words and 9 whole-letter rewrites all scored 100% AI. The writer's raw edit shared about 4% of its 4-grams with the AI draft. The adopted text kept a character similarity of 0.90 to that raw edit. The human letter does not say the same things as the AI drafts: the comparison's writer cut some facts, and this skill does not copy that cut. This is one human letter. It is not a way to pass GPTZero. The public record is the 17 November 2026 letter comparison.

Keep every fact, uncertainty, and qualification. Never invent a fact, an error, or a typo. Do not copy outside text word for word as a technique.

[newer evidence for this letter; older email evidence kept] Stages 28–31 (16–17 November 2026) are the older short-email record. They say to polish a human draft lightly and keep similarity at roughly 0.75 or higher. That 0.75 is 1 − Levenshtein/max_len on public Enron emails, and it was not tested on long documents. It is not this letter's 0.90 character similarity. Both figures stay. For a high-stakes formal letter, this 17 November letter is the newer evidence: a human chooses the words, and AI stays on fact checks and small fixes. Do not read the email light-polish note as permission to polish this kind of letter. Stages 32–34, whose calendar date was not in the supplied record, say to edit the writer's own draft lightly and keep phrasing and structure that already works. They also say that keeping one human sentence word for word is not a technique. Stage 41, noted with those stages on 17 November 2026, added two of the recommender's own public posts verbatim to an AI-drafted recommendation letter and left it at 100% AI. That agrees with the ban on copying outside text. Those email rows stay. None of this lowers the score of a document that starts AI-written. No letter text is recorded.

  1. Have a human choose the words of the whole document. Let AI check facts and fix small errors only. [our scans; n=1] This is the strongest lesson from the 1% versus 100% split, the new wording (about 4% shared 4-grams), and the light-edit budget (0.90 similarity). It matches the existing rule to start from a known human-scoring version and change as little as possible. Stage 16 found that a fresh draft and 11 whole-text rewrites of one synthetic complaint scored 100% AI, including one rewrite built from a previous human-scoring email. A rescan of that earlier email stayed at 3% AI. The lesson does not lower the score of a document that starts AI-written. The cause was not isolated.
  2. Do not treat surface statistics as a humanizing target. Plainer words, a lower type-token ratio, short sentences, varied sentence lengths, fewer nominalisations, and fewer stock phrases are readability advice. They are not detector levers. [counter-evidence, n=1] The human letter was more lexically diverse (type-token ratio 0.608 versus 0.514; rewrites 0.481–0.530), used more long words (7+ letters, 0.334 versus 0.192; rewrites 0.142–0.201), was more nominal (6.0 per 100 words versus 2.3; rewrites 1.8–2.7), and was more uniform in sentence length (SD 4.3 versus 9.1; rewrites 4.8–16.0). It also had more stock evaluative phrases (7 versus 0–2). It scored 1% AI. AI rewrites that pushed any one of these stayed at 100%. [existing evidence, kept] Stage 22 already scored 100% AI on uneven sentence length and on plain verbs with fewer nominalizations. GPTZero's support page says it stopped using perplexity and burstiness in autumn 2023. That vendor claim stays. This letter does not replace it.
  3. Give the title in one clause and point to an enclosed CV. [hypothesis; untested in isolation] Credential facts move into the enclosure. They are not deleted. Credential sentences were the top-flagged sentences in four runs. The human version kept only a title clause, with the CV enclosed. A cheap isolating test, the AI baseline with only the credentials paragraph removed, has not been run. An unrun test is not a finding.
  4. Name the benchmark, tool, or venue for the expert reader. [hypothesis; a tendency, not a rule] Keep a definition to one short sentence. Avoid a step-by-step lay explanation and a rhetorical question. The AI drafts used a tutorial register, a lay step-list was flagged, and on a second letter the 100% chunk was the one explaining an event to an outsider. [counter-evidence] A chunk with a one-line definition scored 7%.
  5. Build each paragraph as a claim, then a dated concrete action by the subject, then a short judgement tied to that action. [hypothesis] This is the human letter's structure. Actor-first sentences on their own did not move the AI drafts: they stayed at 100% even with 4–6 sentences that had the candidate as subject. Use the shape as structure, not as a trick.
  6. Leave conventional professional phrasing in place when an example backs it. [counter-evidence, n=1] Phrases such as "demonstrates leadership" or "gives me confidence" are normal in a human formal letter. The human letter had 7 checklist hits at 1% AI. The AI versions had 0–2 hits at 100%. Backing each phrase with a stated example is a writing check. It is not a measured score lever.
  7. State each needed qualification once, and plainly. [hypothesis; confounded] Do not add a hedge, a disclaimer, or an approximation nobody asked for. The AI drafts carried 3 negations and 3–5 approximators, against 0 and 1 in the human version. Part of that gap is that the comparison's writer cut facts. This skill keeps those facts. A scope or independence statement can sit where the format expects it, such as a capacity line or a note to the signer, and the signed letter still states it when it is needed.
  8. After a human draft, keep later edits small and fact-driven. Fix an overclaim or an overlap. Keep the writer's sentence shapes. Do not run a whole-text polish. [our scans; n=1] The adopted text stayed at 0.90 similarity and scored 1% AI. Its residual "AI Polished" label may reflect the small AI-made fixes. That cause was not shown: the free plan gave no sentence highlights. On a different letter, 16 November 2026, a full natural polish of c7 (15% AI) was a new text, c8, at 100% AI. Neither result is a way to pass GPTZero.
  9. Judge by the whole-document score. Use a chunk scan only to find which part is pulling the score up, then route that part back to a human. [our scans; one second letter] That letter, hand-edited at sentence level with no human writer, scored 53% overall. Three chunks scored 32%, 100%, and 7%. AI rewrites of the 100% chunk scored 100% again and pushed the whole letter back to 100%. A paragraph score is still not the assembled document's score. When a scan was requested, rescan the exact final document.

Establish the purpose and facts

  1. Save the original text. Identify its reader, purpose, format, and requested tone or length from the supplied context. For a discussion reply, also read the question and existing replies. For a private retailer or customer-service email, follow Private retailer and customer-service email. For public participation, respect the destination's rules, explicit objections to AI-assisted work, and disclosure requirements.
  2. Treat supplied facts as the starting point for a self-contained edit. Verify substantive factual corrections against relevant sources; for technical advice, inspect official documentation or implementation and record the version when it matters. Distinguish source inspection, local tests, and deployment verification. Flag unsupported claims instead of adding invented evidence.
  3. Inventory the substantive claims, commands, code, conditions, uncertainty, and reported symptoms. Record each claim's time scope: a dated observation and a limitation that still applies are different facts. For each claim record whether it is preserved, corrected, or removed, and why. Identify what the reader needs to understand, decide, or do.

Private retailer and customer-service email

Use this section for a short private email or message to a retailer, marketplace seller, or customer-service desk. For that email, this section wins every conflict with Write and review. Write and review applies only where this section is silent. Public discussion replies, articles, and documentation stay with Write and review.

Put the facts first. Open with the answer to the latest question, then the other supplied facts and qualifications, then one clear ask. Use ordinary words. Keep every supplied fact and qualification, including limits, dates, conditions, and uncertainty. Concision is wording and order. It does not omit a supplied fact, and the linked-evidence scope cut does not apply to this email. Keep the user's sign-off immediately after the ask, including the closing and name they supplied. When they supplied no sign-off, leave the email without an added name, title, or contact detail.

Work each supplied fact into a sentence that is already there: the answer, the fault, or the ask. Do not add a standalone "Please note..." sentence or a reminder sentence that only lists facts. A bolted-on recap or date sentence, and the sign-off or signature block, are the first places to check when a requested scan flags the draft. On one synthetic kettle email, folding a date into an existing sentence, and avoiding the persona name Helen, removed those two triggers. That is one email. It is not a way to pass a detector. Change how a fact is placed or phrased. Never drop a fact, an uncertainty, or a qualification to lower a score. The sender's own name is a fact: keep the name they supplied, and never replace it with a different name to lower a score.

Later preliminary probes on other synthetic emails (Stages 10–15) found further limits. They are not a way to pass GPTZero. Deleting text from a human-scoring raw draft can flip the score to 100% AI: that happened on the jacket (18% to 100%) and the blender (1% to 100%); the kettle deletion-only control stayed at 4%. On the blender raw, deleting an invented address was the single edit that flipped 2% to 100%, while adding real facts into existing sentences (countertop and date, fault details, a refund line) stayed at 5–10% AI. That address deletion did not flip the jacket raw (29% after deletion, 38% after a fact-neutral replacement). Human-leaning blender scores so far still depended on invented details such as an address or a model number. A fact-complete blender with those invented details removed scored 100% AI, including two later versions where the base model wrote the replacement wording. Never invent an address, a model number, or any other fact to chase a score. Keep every real fact, and never invent one.

[evidence + our scans] When a known human-scoring version of this email exists, start from that version and change as little as possible. Do not run a whole-text AI polish or rewrite pass over fact-dense text. On 13 November 2026, free-plan GPTZero Model 5.0b scored one synthetic pot-handle complaint, and 11 fresh rewrites of it, at 100% AI. The rewrites included plain, passive, mixed, timeline, legal, and casual styles, signed and unsigned, and one built from a previous human-scoring email. A rescan of that earlier email stayed at 3% AI. Whole-text rewriting is not a way to pass GPTZero. The table is the Stage 16 report.

[our scans; mechanism hypothesis] Add a new fact inside a sentence that is already there, for example as a relative clause. Do not add a new fact as a bracketed parenthetical, and do not append it as its own sentence. Change one thing, then scan, before changing anything else. On that complaint, the parenthetical "(bought on 18 October 2026, £42.75 paid with Shop Pay)", which added three facts, scored 100% AI. The photos fact folded into the middle of an existing sentence scored 7% AI (177 words). The same fact as a clause on the first sentence scored 100% AI, and as its own final sentence scored 100% AI. One appended photos sentence on the human-scoring control scored 57% AI. These rows are one email. They do not explain the model, and they are not a way to pass GPTZero.

[our scans] Small edits on other synthetic complaint emails (15–16 November 2026, Model 5.0b Basic, three free accounts shared across the probes, single scans unless replicated) gave no general lever. Brackets around words already in the email changed nothing (four emails, 0 points), so the bracket rule above is about adding facts, not punctuation; use brackets where they read naturally. Sentence-rhythm and punctuation changes, a paragraph split, and removing stock connectives ("Unfortunately", "As such", "therefore") left 100% emails at 100%. Single-word synonym swaps moved human-scoring emails by up to 8 points, and the direction was not stable. Do not swap synonyms to chase a score. Moving the request to the start dropped two emails (100 to 8, and 100 to 54, each replicated once, at 8 and at 54) but left four other 100% emails at 100%, and a natural rewrite of the 8% email with the request first scored 100%. Keep the request last and the English natural. None of this is a way to pass GPTZero. Details: detector research.

[evidence] Never invent a detail, and never delete a fact, an uncertainty, or a qualification, to chase a score. Before any requested scan, run scripts/check_facts.py on the fact-complete source and the candidate. If it reports a missing, changed, or added number, date, price, order ID, or capitalised name, correct the candidate before scanning. The checker does not judge style or a detector score.

[hypothesis] Prefer plain first person, and keep legal boilerplate minimal, where the facts allow. This is a readability preference. The 13 November 2026 scans do not show that plain first person, or shorter legal wording, lowers a GPTZero score.

Write one email to one recipient. Do not prepare a follow-up or a copy for another recipient unless the user asks for that separate email. The one allowed extra wording is a measured alternative the user explicitly requests: keep the original as the control, preserve every supplied fact and qualification, and do not run a scan. It is not a second recipient.

When the supplied draft already answers the latest question, keeps every supplied fact and qualification, uses one clear ask and ordinary words, and keeps the user's sign-off, leave that draft unchanged and say so. After any edit pass, also return the change list and fact checklist in Write and review. Still apply a requested factual correction or a sign-off the user asked to add. Also fold a bolted-on recap or date sentence into a sentence already there, and keep the fact. That fold removed a date-sentence trigger on one kettle email. It is not a way to pass a detector.

Never invent personal experience, an injury, storage or carrying details, a cause, a date, tests, certainty, identity, quotations, typos, slang, deliberate errors, or invisible characters. Keep an unknown fact unknown. When the email cannot proceed without a missing fact, ask for that fact in the one ask.

Do not copy a public asker quotation, or any other quoted line, into private mail to humanize it or to seek a lower score. Do not add, remove, or alter a fact, uncertainty, or qualification to chase a detector score. Change how it is placed or phrased instead. A seller's claim that the text is AI-written, or a report that spam filtering held it, does not ask for a detector scan. It establishes no baseline, model, score, or reason for filtering. Do not promise a human classification, a lower detector score, or inbox delivery. The kettle sentences above limit that one email. They are not a method that passes a detector. A lower score requires a comparable scan the user explicitly requested.

Write and review

  • Separate the source inventory from the answer inventory. For a direct question, keep the conclusion, necessary support, and qualifications that change the reader's decision; leave other background in the linked evidence. Record scope cuts rather than carrying every source fact into the reply.
  • Rebuild the text around the reader's purpose rather than retaining the original sentence order. Put the main point where the format calls for it, connect each sentence to the next, and keep the detail needed to understand or act on it. After moving evidence, check paragraph density: a paragraph that combines the intervention, results, screenshot explanation and limitations may need a break even when every sentence is accurate.
  • For a procedure or causal explanation, map each action to its component, prerequisite, and limiting condition. Arrange the answer in that dependency order. Place a prerequisite before the action that needs it and keep an exception beside the claim it limits. Replace an ambiguous pronoun with the existing component name. Keep unspecified actors unnamed, including in intermediate fact tables: “All 180 articles have been copied” does not establish that a migration team copied them. Preserve passive voice when naming an actor would add an unsupported fact. For ordinary polishing, leave already clear passages intact. When the user explicitly requests a measured alternative, retain the original as a control and allow a different reader-appropriate framing that preserves every supplied fact, qualification, and the practical meaning.
  • Prefer concrete nouns and verbs. Fewer nominalisations and plainer verbs are readability preferences when a sentence is easier to read that way. They are not detector levers. [counter-evidence, n=1] On 17 November 2026 the human letter had 6.0 nominalisations per 100 words and scored 1% AI. AI rewrites at 1.8–2.7 nominalisations per 100 words stayed at 100%. [existing evidence, kept] Stage 22 already scored 100% AI on plain verbs and fewer nominalizations. Remove generic reassurance, repeated summaries, rhetorical contrasts, and irrelevant background. Use lists only where they make steps or comparisons easier to follow. After factual repair, check for stiff noun phrases and overloaded steps; moving unchanged sentences may leave those problems intact. See High-stakes formal writing.
  • For vocabulary cleanup, use the phrase review list. Avoid stock phrases when they add no meaning; review matches in context rather than banning words or swapping synonyms mechanically. Preserve precise terminology, intended tone, uncertainty and protected spans. An absent phrase needs no edit, and a cleaner vocabulary is not evidence of a lower detector score. [evidence] A synonym swap, an "AI word" list, invisible characters, DIPPER, and a commercial humaniser are not detector levers. Use a plainer word when it is easier to read. [counter-evidence, n=1] Plainer vocabulary, a lower type-token ratio, and removing stock evaluative phrases are readability advice. They are not detector levers. On 17 November 2026 the human letter was more lexically diverse (type-token ratio 0.608 versus an AI baseline of 0.514; rewrites 0.481–0.530) and had more stock evaluative phrases (7 versus 0–2) at 1% AI. AI rewrites that pushed the opposite stayed at 100%. Leave conventional professional phrasing when a stated example backs it. Do not scrub it to seek a lower score. Do not swap synonyms, strip a listed word, insert an invisible character, run DIPPER, or run a commercial humaniser to seek a lower score. The 13 November 2026 literature update records why that does not transfer to GPTZero Model 5.0b. A later back-translation experiment does not reopen those levers. The 16 November 2026 literature update restates that limit. [our scans; existing evidence, kept] A third-party humanizer bot rewrites the draft with another model. AI-on-AI rewrites of AI-drafted recommendation letters did not lower the score in this repo: Stages 22–23 scored all 30 natural fact-complete variants at 100% AI on free-plan GPTZero Model 5.0b Basic, including a pstack unslop rewrite and a pstack technical-writing rewrite (evidence limits). That Stages 22–23 record stays. The 17 November letter is newer, and its own AI rewrites also stayed at 100%.
  • [evidence] During a requested detector iteration, rescan the exact final assembled document. A paragraph score is not that document's score. Treat a fluency polish of a low-scoring draft as a new text and rescan those bytes. When readability and a lower score conflict, report both and keep natural, fact-complete English. Round-trip translation is an experiment only when the user has asked for detector iteration, not a default rewrite. Repair only factual drift and genuine grammar errors, fact-check, and warn that odd phrasing may be doing the work. Do not assume a pivot language. Fixing the worst phrases one at a time and rescanning is one observed sequence on a single letter. It is not a bypass and not a recipe. The measured case is one private recommendation letter on 16 November 2026, free-plan GPTZero Model 5.0b: c4 100% AI, c7 15%, c7a 13%, c7b 43%, c8 100%. The user chose the natural c8. The limits are in detector research. This is one letter, one account, and a single scan of each version. It is not a way to pass GPTZero. A chunk scan only finds which part is pulling the score up. Route that part back to a human, then judge the whole document. [our scans; one second letter] On 17 November 2026 a hand-edited letter scored 53% AI overall, with chunks at 32%, 100%, and 7%. AI rewrites of the 100% chunk scored 100% and pushed the whole letter back to 100%. See High-stakes formal writing.
  • Preserve operational qualifications and exact code unless evidence supports a correction. Record deliberate scope cuts; do not call an answer fully claim-preserving when useful details were removed.
  • Polish the explanatory message independently of its code. Keep code blocks and standalone commands fixed during prose experiments; assess their correctness separately, and record any necessary technical correction. Keep inline identifiers and values where they are part of an explanation. The full message, including its code, still needs a consistency review.
  • Match the requested register. If an authentic writing sample is supplied, identify its register, paragraph openings, and transitions before drafting; preserve useful habits during cleanup without importing the sample's facts, opinions, or personal history. Vary sentence length when it improves pacing. Do not vary it to seek a lower score. [counter-evidence, n=1] Sentence-length variety, including the old burstiness target, is readability advice. It is not a detector lever. On 17 November 2026 the human letter's sentence-length SD was 4.3, more uniform than the AI baseline of 9.1, and the letter scored 1% AI. AI rewrites with SDs from 4.8 to 16.0 stayed at 100%. [existing evidence, kept] Stage 22 already scored 100% AI on uneven sentence length. GPTZero says it stopped using burstiness in autumn 2023. That vendor claim stays. Do not force informality or strip useful structure. Never invent personal experience, injuries, storage or carrying details, causes, dates, tests, certainty, identity, quotations, typos, slang, deliberate errors, or invisible characters to make text appear human-written.
  • A short, exact asker quotation can identify the concern being answered in a public discussion reply. Include it only when it helps that reader. Verify context and attribution; do not pad or rotate quotations to hunt for a detector score. Do not import that technique into private mail to humanize it.
  • Compare the finished answer directly with the claim inventory. Check who acts, negation, cause and effect, scope, certainty, numbers, units, and prerequisites. Keep an ongoing limitation current; changing “has not been validated” to “had not been validated at that point” narrows the claim unless the evidence supports that change. Match protected code, commands, identifiers, URLs, and exact quotations byte-for-byte unless a verified correction is recorded. Account for every cut or correction and remove unsupported additions. This writer self-check does not replace separate review.
  • When a separate reviewer is requested or available, give them the reader's purpose, relevant context and sources, baseline, final text, and claim inventory without detector scores. Request a verdict on correctness, relevance, readability, tone, and necessary qualifications, tied to the final file's SHA-256. Resolve factual objections even if an incorrect version has a lower score. If no separate reviewer is available, say so; a self-check is not independent review.
  • [our scans; hedged] When a human-written draft already exists, polish it lightly and keep similarity to that draft at roughly 0.75 or higher. That figure is a conservative margin on this sample, not a tested cutoff. Similarity is 1 − Levenshtein/maxlen: Levenshtein is the character edit distance, and maxlen is the length of the longer text. This repo does not include a script for that ratio. Check facts against the draft. On free-plan GPTZero Model 5.0b Basic scans of public Enron AESLC test-split emails (Stages 28–31), LLM edits of human drafts at similarity of at least 0.70 stayed classed human in 8/8 recorded edits (a1, a2, a3, b1, b2, f1, g1, h1). That 8/8 is eight edits, not eight drafts. Seven of those eight were 0–6% AI. g1 rose to 21% (+19), past the about 2-point rescan band, on the assessment/judgement-style draft; its similarity was 0.751. This only preserved a human classification when the starting draft was already human-written. It does not lower the score of text that starts AI-written, and it does not guarantee 0%. Keeping the draft's casual voice was not supported (a4, similarity 0.502, scored 100% AI). The committed rows name 7 drafts (AESLC test rows 50, 57, 65, 79, 101, 119, and 170). An eighth draft is not in those rows and is not counted. Lengths were not logged. There was one upward drift (g1), and it was not tested on long documents. This note does not replace the rules above, and it is not a way to pass GPTZero. Newer evidence for a high-stakes formal letter is the 17 November 2026 section above. That letter's 0.90 figure is character similarity on one letter, not this Levenshtein margin. Both stay. For that letter, a human chooses the words, and this email note does not license a whole-text polish. Details: detector research and the GPTZero results report.
  • [our scans; hedged] When the starting point is the writer's own draft, edit lightly and keep their phrasing and structure where it already works, rather than polishing every sentence. Heavy, uniform polishing of a whole text is what tends to get flagged. That reading is from a few short public Enron AESLC test-split emails on free-plan GPTZero Model 5.0b (Stages 32–34): one detector version, and mixed scan modes as already recorded. Stage 32's scan mode was not in the supplied record. Most Stage 33 scans were Basic after the free Advanced allowance ran out, and which of the six used a remaining Advanced scan was not recorded. Stage 34 rescanned the unchanged heavy rewrites in pure Basic mode at 100, 24, and 14. This applies only when the draft is the writer's own. It does not lower the score of text that starts AI-written, and it does not guarantee 0%. Keeping one human-draft sentence word for word, or pasting outside human text into the document, is not a technique in this skill. It does not replace the light-edit note above, and it is not a way to pass GPTZero. For a high-stakes formal letter, the 17 November 2026 section above is the newer recommendation: a human chooses the words, and AI stays on fact checks and small fixes. This email note stays beside it. Stage 41, recorded with these stages, left an AI-drafted recommendation letter at 100% AI after two verbatim public posts, which agrees with that section. Details: detector research and the GPTZero results report.
  • After any edit pass, return a short change list the human can use to verify fact preservation and to keep the edits small. For each change give the original phrase, the new phrase, and the reason. Then confirm a fact checklist: every fact, number, date, uncertainty, and qualification from the source is still present. When the starting text is a human draft, the list shows that the edit stayed small and fact-driven. If nothing changed, say so and still give the checklist. The list is for that review. It is not a detector result.

Measure only when requested

Read measurement instructions before live scanning and evidence limits before interpreting results. Use the user's requested service, mode, and supported browser tools. This skill supplies no browser controls and does not authorize subscriptions, paid scans, public posting, or new account access.

Run a scan only when the user explicitly asks for a detector measurement. A seller's claim that a message is AI-written, or a report that spam filtering held it, is not that request and does not establish a baseline, a displayed model, a score, or why a filter acted. Do not promise that any revision will be classified as human-written or delivered to an inbox.

When explaining detector mechanisms or planning another batch after unchanged results, read detector research.

When the user explicitly asks for an autoresearch workflow to improve this skill, read the bounded research workflow. It separates skill proposals, preference-based selection, quality review and measured confirmation; ordinary writing does not need this workflow.

For a comparison, select cases and candidates before seeing scores. Record fresh baseline and candidate scans using the same service, mode, and displayed model. Retain failures and unchanged results. Compare a fixed body with a relevant quotation separately from prose rewriting; track content and length changes. Distinguish first-pass trials from adaptive iterations. Record an unchanged candidate as a no-op. Comparing it with a separately rewritten baseline measures preservation versus that rewrite, not a benefit from edits the candidate did not make.

When a scan runs, the detector target is the message's prose unless the user explicitly requests the whole message. Declare prose_only before testing, exclude only independently identified code blocks or standalone command lines, and retain all explanatory prose, qualifications and inline identifiers. Save the full reviewed message, exact prose input and code-range mapping separately. Use the same scope for both arms; keep earlier whole-message scores under their original scope. An explicit whole-message request uses full_text instead.

Confirm the effective editor text, completed scan, fresh-result indicator, model, native result labels and metrics, timestamp, and warnings. Save exact inputs and private visible evidence. A stale displayed number is not a result for changed text. Never pad short prose, add code back merely to reach a detector minimum, or drop a fact, an uncertainty, a qualification, or a difficult explanatory passage to lower a score. Change how a fact is placed or phrased instead.

When a requested scan flags a short email, check a bolted-on recap or date sentence and the sign-off or signature block before rewriting the rest. To locate a trigger sentence, follow one-fact restoration in the measurement instructions: scan a deletions-only control, then variants that each add one fact back. The evidence limits keep the kettle scans preliminary. On that one email, folding the dates into an existing sentence removed the standalone date-sentence trigger, and the persona name Helen scored 100% AI where the sign-off Sue Hall scored 13% AI. A stock-persona-name signal is a hypothesis, not a finding. A Basic rescan of the Sue Hall text reproduced 13% AI. Three other synthetic emails built the same way, with plain sender names, each scored 100% AI. Date folding and avoiding a stock persona name removed triggers on the kettle email. They are not a way to pass GPTZero. The control is missing facts, so it is not the email to send, and a real sender's name stays as supplied. The later blender and jacket probes are also preliminary: deletions can flip a human-scoring raw draft to 100% AI, adding real facts into existing sentences was mostly safe on the blender, deleting an address did not flip the jacket, and the human-leaning blender scores relied on invented details that must never be used. See the GPTZero results report.

[evidence] Before any requested scan, run scripts/check_facts.py on the fact-complete source and the candidate. Scan only a candidate that preserves every extracted number, date, price, order ID, and capitalised name. Every scan session includes a known-human control and a known-AI control, on the same service, mode, and displayed model. Log the model version, whether the scan was Basic or Advanced, any use of a free account's complimentary Advanced Scan, the word count, the credits, and a "Possible AI Paraphrasing" label (including any Generated-versus-Paraphrased split) as its own field. Do not treat a Basic result and an Advanced result as a pure wording effect until the same bytes are scanned in both modes. A session without those controls is not evidence about a wording change. The Stage 16 pot-handle scans are one email on Model 5.0b. They are not a way to pass GPTZero.

Assembled documents, polish, and round trips. The rules below are from one private professional recommendation letter, about 700 words, not published. Free-plan GPTZero Model 5.0b, 16 November 2026. Word counts are wc counts. Fact checks used a fixed 82-item inventory. The full note is the 16 November 2026 literature update. This is one letter and one account, a single scan of each version. It is not a way to pass GPTZero.

[evidence] Document scores and paragraph scores are different quantities. After any paragraph-level experiment, rescan the exact assembled body before reporting a letter or article score. A 0% paragraph does not establish a low document score. On this letter one round-trip paragraph (v4, 120 words) scored 0% AI, and that paragraph result replicated on a rescan. The assembled letter c4 (698 words) scored 100% AI and was labelled "Possible AI Paraphrasing". GPTZero's Advanced Scan page says the percentage is confidence about the entire text, and that the model reads the document as a whole. That page is a vendor claim. It matches the c4 split. It is not a causal editing rule.

[evidence] A fluency polish of a low-scoring draft is a new text. Rescan the exact polished bytes. Do not treat a "natural academic English" pass as cleanup of an already scored candidate. On this letter, c7, c7a, c7b, and c8 were each logged once on the same fresh account, Model 5.0b, Advanced. c7 (assembled round-trip paragraphs, 702 words, 82 facts still present) scored 15% AI, with the label quoted as highly confident that the text is entirely human. c8 (full natural polish of c7, 672 words, 25 targeted replacements, 82 facts still present) scored 100% AI. Earlier full-letter and most paragraph scans of this letter used Basic. Basic and Advanced were not paired on the same bytes, so that mode gap is a recorded caveat, not a measured cause. The free-plan page describes one complimentary Advanced Scan. Four Advanced logs on one account exceed that description, so the mode label is the session record.

[evidence for this letter] Partial repairs were scanned as their own texts. c7a (c7 plus the five most awkward phrases fixed, 689 words) scored 13% AI, highly confident human. That 13% is not a step up from c7's 15%. c7b (c7a plus eight further outright-error fixes, remaining stylistic quirks left in, 676 words) scored 43% AI, uncertain, likely human. Whether the translation-like phrasing left in c7 caused the 15% is a hypothesis. The sequence is one letter. It is not a bypass and not a repair recipe. Each step is a new text.

[evidence for this letter; literature as supporting, not 5.0b] When readability and a lower score pull in opposite directions, report both. Do not optimise the score. Keep natural, fact-complete English as the text to use. Keep the awkward low-scoring draft as a scored backup. On this letter the user chose c8, because quality beats the score. c7 remains a scored backup. TH-Bench's abstract, in detector research, is supporting literature on open detectors.

[evidence that a round trip can move a paragraph; evidence that it need not move a 5.0b document] Round-trip or back-translation is an experiment, not a default rewrite. Use it only when the user has asked for detector iteration. Repair only factual drift and genuine grammar errors. Keep every protected name, number, date, hedge, and highlighted note. Fact-check before scanning. Say plainly that odd phrasing may be part of any lower score. ESPERANTO measured an older GPTZero, a different pipeline, and true-positive rate rather than a 5.0b AI percent. That literature does not make a round trip a method. Full-letter c4 was caught as possible AI paraphrasing.

[evidence of the contrast on this letter; mechanism is a hypothesis] Do not assume a pivot language. On this letter, German round trips stayed fluent, kept English word order, and scored 100% at paragraph level. Some Chinese round trips restructured the sentence and scored low. Japanese did not salvage the results-heavy project paragraph, which failed in eight versions at 94–100% AI. ESPERANTO Table 1 shows German back-translations staying closer to the English source than Chinese or Japanese ones. That contrast is compatible. It is not a 5.0b mechanism proof.

[evidence for this letter only] Folding facts from a failing paragraph into a neighbouring passing paragraph is how c7 was built. It is one measured assembly, not a recipe. It still requires a full-document rescan, a fact check, and a readability review.

[evidence of a mode caveat; vendor pages for what the fields mean] Log scan mode, model version, and the paraphrase label separately from the AI percent. Record Basic or Advanced, and whether a free account's complimentary Advanced Scan was used. Record "Possible AI Paraphrasing", and any Generated-versus-Paraphrased split, as its own field. A score is still not authorship, inbox delivery, or permission to send.

[literature and vendor claims; not a 5.0b false-positive rate] Park, Jeong, and Kim, and Li and Wan, show that open detectors can respond to proficiency and to professional academic editing. Garland argues a score should not be the sole evidence in a misconduct decision. GPTZero's Advanced Scan page and its "How AI detectors work" page say a detection result should not be the only basis for punishment, and that the percentage is not the share of the document written by AI. Those sources do not transfer their rates onto Model 5.0b.

[already covered] Do not add DIPPER, a commercial humaniser, synonym lists, or invisible characters as score levers. The 13 November note already covers them, and 5.0b's vendor notes target paraphrased and polished text. A back-translation update does not reopen them.

Proposed experiments, not yet run. Rescan the best text twice plus the known-human control, on the same model and mode, before treating a small gap as a real move. (The brackets and synonym-swap comparisons have since been run on synthetic emails; see detector research.) On the 16 November letter, still not done: pairwise Basic versus Advanced rescans of the exact c7 and c8 bytes, and a second independent rescan of c7, c7a, or c7b on the same mode before treating a single Advanced scan as stable. The c7a 13% and c7b 43% scores above are recorded single scans. An unrun experiment is not a finding.

The optional checker in scripts/validate_receipt.py checks strict plain-text receipt consistency. Its receipt format rejects unrecorded editor changes; it does not support rendered-Markdown transformations or authenticate the service. A valid receipt can record 100% AI. Apply a numeric threshold only when the user explicitly requests one.

Publish and report

Honor existing user authorization; do not invent authorization or ask again for an already authorized action. When the text is a private retailer or customer-service email, return that one email to the user, do not post it, and do not send it unless the user has explicitly authorized the send. For a public reply, refresh the discussion before posting and reassess if context changed. Match the full final body to the reviewed version and the declared detector input to its scan and requested threshold. A prose-only result describes the prose, not the complete message or its code. Under prose_only, changing only code requires renewed full-message review and an updated prose mapping; if the measured prose bytes stay identical, its receipt still describes those bytes. Under full_text, a code change also requires a new scan. Update the owned answer when editing, then independently read back its author, exact body, and URL. Do not post a duplicate or claim an answer was accepted without checking.

Report the actual quality findings, scores and warnings when measured, and edit or keep decisions. When a natural, fact-complete wording and a lower detector score conflict, report both. The text to use is the natural wording. Keep the low-scoring draft as a scored backup. On 16 November 2026, for one private recommendation letter, the user chose the natural version (c8, 100% AI) over the awkward draft (c7, 15% AI). Missing required review or measurement leaves a candidate unpublished. A detector threshold applies only if requested. A successful local receipt check never authorizes publication.

For a showcase, include only quality-passing candidates with a measured decrease against a saved before-text using comparable completed scans. Link exact inputs, identify the intervention and tested version, and keep the full comparison denominator visible. Put unchanged, worse, rejected, unpaired, and unmeasured outcomes in a separate limitations section while retaining all records. A quotation effect is not a rewriting effect, and an untested skill revision is not an established improvement. If a case does not improve, document when the approach may not help or test a specific new edit within a bounded experiment.