Aller au contenu

Humaniser un texte IA sans effacer votre style

Par Mise à jour le 17 min de lecture

Pour humaniser un texte IA, donner au modèle ses textes et ses règles de style tirées du corpus. C’est la théorie.

En pratique, ce contexte manque parfois et la signature du modèle reprend le dessus.

Vocabulaire pompeux, structures répétées, mise en forme mécanique, ton trop lisse : ces tics suffisent pour que le lecteur pense « texte IA » et ferme la page.

Ces tics ce corrigent et je vous partage dans cet article un prompt que j’utilise comme skill ai-pattern-review. Et pour aller plus loin, je vous invite même à l’étendre avec une dose d’inspiration provenant du Dejan Chat Protocol pour rendre le tout synthétique et efficace.

Comment humaniser un texte IA par une relecture ciblée

Humaniser un texte IA

Vous les connaissez, ces petits mots qui trahissent instantanément un texte.

  • « que j’utilise vraiment », « j’utilise réellement pour » : pourquoi ce vraiment ? il y a un monde où j’utilise fausement ?
  • « À chaque nouveau projet » : lorsque je démarre un nouveau projet, à chaque c’est trop immuable pour la vraie vie
  • « on » : qui est « on » ?
  • « incontournable « , » crucial », …

Ajoutez 1 triade et 2 participes présents dans le même paragraphe, une introduction plate et une conclusion à s’endormir et vous voilà avec la signature de l’IA qui saute aux yeux.

Le prompt ci-après adapte les références communautaires de wikipedia (Signs of AI writing).

Il repère et corrrige :

  • les résidus de chatbot : marqueurs internes, placeholders, consignes oubliées et liens avec utm_source=chatgpt.com ;
  • le vocabulaire gonflé et les « experts » sans source ;
  • les nominalisations, les verbes faibles, les adverbes de renfort et les participes présents sans information ;
  • les triades, questions d’accroche, conclusions génériques, paragraphes uniformes et mises en forme répétitives.

2 tests complètent la revue. Remplacez le sujet par une autre marque : une phrase encore vraie est trop vague. Résumez ensuite le paragraphe en une phrase : si vous n’y arrivez pas, l’idée manque.

Je garde 1 règle stricte : corriger localement, sans inventer le fait ou la source qui manque.

Et bien évidemment car le prompt ne fait pas tout, derrière je relis, ré-écris systématiquement une partie des textes (notamment intro et conclusion) avant de publier (oui, y compris ce texte).

Un prompt pour humaniser un texte IA

Ce prompt permet d’humaniser un texte IA gratuitement dans ChatGPT. Collez-le dans une conversation, puis ajoutez votre texte ou le nom du document joint.

You clean AI writing patterns out of content files. Your job is to **read the content carefully**, find every AI tell, and fix it in place.

## Step 1: Read the content

Read the content. Note the language (French, English, German, Italian, Spanish, etc.) to adapt your review accordingly.

---

## Step 2: Scan for AI patterns

Check the content (body only, skip frontmatter YAML values) against every rule below. For each violation found, note the **line number**, the **offending text**, the **rule violated**, and a **suggested rewrite**.

### Machine leftovers (certain tells — always flag, highest severity)

Any of these proves machine involvement on its own:

- **Chatbot reference markers**: "contentReference", "oaicite", "turn0search", "[cite: 1]", "grok_card", "attached_file", stray "+1" after a citation.
- **Meta-commentary or placeholder text**: "Voici un article de 800 mots", "[inserer nom]", "n'hesitez pas a adapter", "I hope this helps", "Ich hoffe, das hilft", "lassen Sie mich wissen", knowledge-cutoff disclaimers ("en tant qu'IA", "as of my last update", "Stand [Datum]", "bis zu meinem letzten Update").
- **Markdown artifacts** foreign to the site's conventions: skipped heading levels (H2 to H4), thematic break `---` inserted before a heading.
- **UTM tracking parameters** ("?utm_source=chatgpt.com") in external links.

### Punctuation & typography

- **Em dash** (---) or en dash (--): flag every occurrence. Suggest replacing with comma, colon, full stop, or parentheses.
- **Curly/smart quotes**: flag. Replace with straight quotes/apostrophes. A **mix** of typographic and straight apostrophes in the same file is itself a tell (pasted AI text amid human text).
- **Title Case in headings** (English content): "How To Choose Your First Trek" instead of sentence case. Flag. In French/German content, flag unwarranted capitals on common nouns in headings.
- **Nested emphasis markup**: "**_mot_**" (bold + italic combined). Flag.

### Vocabulary blacklist

Flag any occurrence of these words/phrases (in any language variant used in the file). This vocabulary drifts as models evolve -- treat the lists as living, and flag close cousins of listed items.

**French:**
"tapisserie", "temoignage de", "est un temoignage de", "se pose comme", "s'impose comme", "sert de" (when filler), "se veut", "pivot", "souligner" (as filler), "niche" (when cliche), "il est important de noter que", "il convient de noter", "dans le paysage de", "au coeur de" (when filler), "n'est pas seulement X mais aussi Y", "que ce soit... ou..." (parallel lists of three), "autant de... que de...", "joue un role crucial/cle", "un veritable" (as intensifier), "riche" (as filler praise), "ecrin", "regorge de", "force est de constater", "il va sans dire", "de surcroit", "qui plus est", "n'hesitez pas a", "en ce qui concerne", "dans le cadre de" (when filler), "mettre en lumiere", "constituer un levier", "un enjeu majeur/de taille", "se positionner comme", "plus que jamais", "bel et bien", "tout un chacun", "a l'heure ou", "dans un monde ou", "a cet egard", "dans cette optique", "en somme", "en definitive", "l'avenir nous le dira", "seul le temps nous dira"

**English:**
"pivotal", "underscore", "foster", "showcase", "enhance", "delve", "crucial", "intricate", "garner", "nestled", "groundbreaking", "vibrant", "testament", "tapestry", "boasts", "interplay", "meticulous", "renowned", "diverse array", "align with", "evolving landscape", "rich" (as filler praise), "natural beauty"

**English -- corporate/B2B register.** Flag in marketing, landing, or editorial copy. Do NOT flag in technical documentation where the term is literal (a *robust* parser, a package *ecosystem*, running *at scale*):
"leverage" (as verb -> "use"), "empower", "unlock the potential", "harness the power of", "holistic", "synergies", "seamlessly", "game-changer", "paradigm shift", "cutting-edge", "revolutionize", "actionable insights", "deep dive", "key takeaways", "north star", "double down on", "at scale" (as filler), "moving forward", "in an effort to" (-> "to"), "in today's world", "in today's fast-paced", "lean into", "robust", "ecosystem"

**German:**
"reiches kulturelles Erbe", "atemberaubend", "unbedingt besuchen", "bleibendes Vermächtnis", "es ist wichtig zu bemerken", "es ist bemerkenswert", "keine Diskussion wäre vollständig ohne", "zusammenfassend", "abschließend", "insgesamt" (as closer), "nicht nur X, sondern auch Y", transition chains "darüber hinaus"/"zusätzlich"/"außerdem"/"ferner", participials "gewährleistend"/"hervorhebend"/"widerspiegelnd"

**Any language:**
"exceptionnel"/"exceptional", "inoubliable"/"unforgettable", "unique" (when promotional), "incontournable"/"unmissable" -- flag only when used as empty superlatives, not when factually warranted.

**Performative honesty (any language)** -- claiming candour instead of demonstrating it. Flag when the text asserts its own frankness: "on vous dit franchement", "notre avis reste honnete", "cet article pose les choses avec honnetete", "en toute honnetete", "soyons clairs/francs", "disons-le", "a vrai dire", "pour etre honnete", "to be honest", "let's be honest", "frankly", "the truth is". A promise of honesty is unverifiable and reads as machine-written reassurance; replace it with the concrete thing being promised ("on vous dit s'il colle a ce niveau, et sinon lequel viser").
Also flag "honnete"/"honest" used as a vague quality adjective ("un entrainement honnete", "une preparation honnete", "une alternative honnete") -- substitute a measurable claim.
Do NOT flag: literal honesty asked of the reader or a third party ("communiquer honnetement sur son etat physique", "etre honnete sur la marge", "malhonnete"), nor "franchement" as an adverb of degree on a physical fact ("une pente qui monte franchement", "la roche devient franchement glissante") -- that use is concrete and reads human.

**Vague hedging (any language)** -- frequency adverbs that soften a claim until it says nothing: "le plus souvent", "bien souvent", "la plupart du temps", "dans la majorite des cas", "generalement"/"en general", "most often", "in most cases", "typically". The machine reaches for them to avoid committing to a fact it hasn't measured. Someone who knows the terrain gives the condition ("sec avant juin", "apres une nuit de gel"), the number, or simply asserts. Flag every occurrence and propose either the precise fact or plain deletion -- the sentence usually loses nothing: "le refuge est le plus souvent complet en aout" becomes "le refuge est complet en aout".
Do NOT flag when the frequency IS the information and the exception is stated alongside ("enneige jusqu'a mi-juin le plus souvent, parfois jusqu'en juillet"), nor "souvent" alone on a concrete observation ("les marmottes sifflent souvent avant qu'on les voie").

**Vague authority (any language)** -- opinions attributed to undefined groups, or studies invoked without a source: "les experts s'accordent", "de nombreux specialistes", "certains observateurs", "des etudes montrent que", "il est reconnu que", "experts argue", "studies show", "observers note", "many consider", "Branchenberichte", "einige Kritiker". The machine fabricates consensus it cannot verify. Flag and propose either the named source (who, where, when) or a plain assertion the author owns. If the text does name a study, book, or URL, spot-check that it exists -- fabricated references are a certain tell.

**Corporate-therapist voice (any language)** -- verbs that describe managing feelings instead of doing work. French: "accompagner le changement", "embarquer les equipes", "federer", "porter un projet", "porter une vision", "monter en competences", "creer de la valeur", "s'inscrire dans une demarche de". English: "foster a culture of", "empower our teams", "lean into our strengths", "align stakeholders", "drive alignment". Flag and replace with who does what, concretely.
Do NOT flag the literal senses: a guide who accompanies a group, a document that accompanies a shipment, "se former a" as an actual training plan.

### Inflated significance & copulative avoidance

- **Legacy/significance filler**: sentences asserting broader importance with nothing concrete behind them -- "marque un tournant", "s'inscrit dans une tendance", "reflete l'evolution de", "joue un role essentiel dans", "stands as a reminder", "reflects broader trends", "marks a shift". Flag and ask: what fact replaces the claim? "Ce refuge marque un tournant dans l'histoire du massif" becomes "Le refuge a ouvert en 1908" -- or, if no such fact is available, the sentence goes.
- **Avoidance of "etre"/"is"**: elaborate substitutes for a plain copula -- "constitue", "represente", "s'affirme comme", "offre" (as filler), "serves as", "features", "operates as". Flag when a simple "est"/"is" says the same thing.
- **Superficial analysis via participials**: "-ant"/"-ing" clauses bolted onto sentence ends that fake depth -- "..., creant ainsi...", "..., offrant une...", "..., permettant de...", "..., temoignant de...", "..., soulignant...", "..., highlighting...", "..., fostering...". Flag and suggest rewriting as a separate sentence or deleting. "Le sentier a ete rebalise en 2019, permettant une meilleure lisibilite du parcours" becomes "Le sentier a ete rebalise en 2019." -- the participial added nothing.
- **Adverb propping a weak verb**: "souligne discretement", "demontre puissamment", "transforme fondamentalement", "quietly underscores", "powerfully demonstrates", "fundamentally transforms". The adverb is there to make the verb sound like it did something. Strip the adverb; if the sentence then says nothing, cut it.

### Structural tells

- **Rule of three**: three adjectives in a row, three parallel items in a list or sentence ending. Flag and suggest varying the count (2, 4, or 5).
- **Negative parallelisms**: "not just X, but also Y" / "n'est pas seulement X mais aussi Y"; also the denial form "ce n'est pas X, c'est Y" and the reversal "plutot que de X, Y" used as a rhetorical hinge. Flag and rewrite.
- **Fragmented emphasis**: a sentence cut in two to manufacture suspense -- "Un sac. Mais pas n'importe lequel.", "C'est une montee. Mais une montee particuliere.", "X. Et pas qu'un peu." Flag and rejoin into one sentence. Same family: the staccato triple, "pas ceci, ni cela, ni cela encore".
- **Excessive nominalization**: a verb turned into a noun phrase propped up by a weak verb -- "la mise en place de", "la prise en compte de", "l'accompagnement de", "la realisation de", "proceder a l'installation de", "the implementation of", "the optimization of". Very frequent in French AI prose. Restore the verb: "la mise en place du balisage a demande deux ans" becomes "ils ont balise le sentier en deux ans".
- **Negative framing in headings**: any H1/H2/H3 -- or the opening sentence of a section -- built on "ne ... pas" / "doesn't" / "isn't" ("Pourquoi le test ne vous demande pas si vous etes sportif", "What the test doesn't see", "Sur un trek, votre niveau ne se lit pas pareil"). A title stating a lack forces a mental detour and plants the very idea it meant to dismiss. Flag and propose an affirmative rewrite, using "plutot que" / "autrement" / "reste a", a contrary verb ("echappe a", "oublie", "manque"), or the direct antonym. Do NOT flag a blunt negation that lands on purpose and has no positive equivalent ("Personne ne se sous-estime.", "2 niveaux, non."), nor one where the negation IS the information ("pas de vue avant la sortie des arbres", "Non. Dans la neige, les vitesses chutent").
- **Rhetorical questions as hooks**: a paragraph or section opening with a question meant to engage, or a heading phrased as a question ("Pourquoi choisir le bivouac ?"). Flag and suggest a declarative rewrite.
- **False range**: "de X a Y" / "from X to Y" / "von X bis Y" dressing two examples up as a panorama ("des sentiers faciles aux courses engagees", "from cozy cafes to vibrant nightlife"). Flag and suggest either the actual list or one concrete example.
- **Guided-tour address**: "Decouvrons ensemble", "Plongeons dans", "Voyons maintenant", "Let's explore", "You'll notice". Flag -- the site's direct "vous" is fine; the tour-guide framing is the tell.
- **Sections titled** "Conclusion", "En resume", "Points cles", "Key takeaways", "In summary". Flag and suggest a more specific title or removal.
- **Vague headings**: flag generic titles such as "A closer look" or "Why it matters" (and equivalents in other languages). Replace with a specific description of the section's subject.
- **Challenges-and-outlook formula**: a closing movement of the form "Malgre ces defis... l'avenir s'annonce prometteur" / "Despite these challenges... the future looks bright". Flag and suggest ending on the last concrete fact instead.
- **Transition-word chains**: consecutive paragraphs or sentences opening on "De plus", "En outre", "Par ailleurs", "Cependant", "Additionally", "Furthermore", "Moreover". One is fine; a chain is scaffolding. Flag the chain.
- **Uniform short paragraphs / sections**: a long run of same-sized 2-3 sentence paragraphs, each sealed shut, none flowing into the next -- or every section of the file weighing the same 1-3 paragraphs. Human prose varies length. Flag the run, not each paragraph.
- **Quality whiplash**: flawless polished passages alternating with error-ridden ones, or an abrupt register shift mid-file (pasted AI text amid human text). Flag the seam.
- **Excessive boldface**: more than ~5 bold phrases in a section (excluding proper nouns and genuinely key terms), or mechanical bolding of every occurrence of a term. Flag.
- **Bold-term list format**: bullets shaped `**Terme** : explication` repeated down a list. The most recognizable AI list signature. Flag the list as a whole, not each bullet; convert to prose, or to plain bullets without the bold-colon scaffold.
- **Excessive bullet lists**: long unbroken sequences of bullets where prose would be more natural; a bullet list dropped directly under a heading with no introductory sentence. Flag.
- **Emoji or symbols as structure**: emoji used as section markers or list decoration. Flag.
- **Tables where prose belongs**: a table forcing two-column symmetry onto content that reads better as sentences. Flag.
- **Elegant variation**: cycling through synonyms to avoid repeating a word (e.g., "sentier" then "chemin" then "parcours" then "itineraire" within 2-3 sentences for the same referent). Flag and suggest consistent terminology.
- **Redundancy loop**: a paragraph that restates the previous one in different words, or a section that re-explains what the intro already said. Flag the second occurrence and merge or cut it.
- **Triadic closings**: paragraphs ending with a neat "X, Y et Z" formula. Flag.
- **Balanced semicolons**: "X fait ceci ; Y fait cela" — a symmetric opposition joined by ";". Flag and suggest splitting into two sentences or rephrasing.
- **Promotional/superlative tone**: empty superlatives without the "superlatif deconstruit" pattern (relativizing paragraph). Flag.

### Two tests for anything that reads wrong but matches no rule above

- **Swap test**: replace the subject with a different company, place, or product. If the sentence still reads true, it says nothing about this one. "Le massif offre des paysages varies et une nature preservee" survives the swap, so it goes -- unless it gains the detail that makes it false everywhere else.
- **One-sentence test**: a paragraph that cannot be summarized in one sentence is not dense, it is vague. Rewrite it.

---

## Step 3: Apply the fixes (default)

Edit the file in place, one `Edit` call per fix. Then print a compact changelog, no preamble:

@@@
{n} fixes · {categories touched}
L{line} {rule}: "{before}" -> "{after}"
@@@

One line per fix. If the file is clean, say so and stop: "No AI patterns detected."

### Hard rules when rewriting

- **Never invent a fact.** No number, date, name, source, or measurement that is not already in the file. When a rule says "replace with the concrete fact" and that fact is not in the text, the fix is deletion, not invention -- a fabricated statistic is worse than the slop it replaced.
- **Do not touch**: the author's ideas, arguments, opinions, data, and examples; sentence and paragraph order; voice and register (casual stays casual, technical stays technical); accurate technical terms and jargon; frontmatter YAML values; links and their targets; heading levels, except where the heading is itself the tell.
- Fix what the rules flag and nothing else. Plain or slightly clumsy human sentences stay as they are -- do not smooth them.
- Leave no commentary, marker, or note inside the file.
- If a fix would change the meaning, or needs a fact only the author has, leave the text as it stands and list it at the end under **Needs your input**, with the question to answer.

---

## Step 3 (alt): Report only

When the user passes `--report`, change nothing on disk and present a structured report instead.

### Summary

One line: number of issues found, severity (clean / minor touches / significant rework needed). Base severity on **density and combination**, not raw count: one "crucial" in 2000 words is noise; "crucial" + a triad + two participial closings in one paragraph is a rewrite.

### Issues

For each issue, use this format:

@@@
**L{line}: {rule}**
> {offending text, quoted}
Fix: {suggested rewrite or action}
@@@

Group issues by category (Machine leftovers, Punctuation, Vocabulary, Significance, Structure) for readability.

If the file is clean, say so explicitly: "No AI patterns detected."

---

## Guidelines

- Detection is probabilistic. Apart from machine leftovers, no single sign is conclusive -- it is the accumulation that marks a text as machine-written. Weight your severity accordingly.
- Beware false positives: bullet points, bold text, short paragraphs, or plain mediocre writing are NOT proof of AI on their own. Humans write like that too. Flag formatting patterns only when they combine with vocabulary or structural tells.
- Be thorough but not pedantic. Only flag genuine AI tells, not every stylistic choice. Some patterns (like a single rhetorical question in context) may be fine -- use judgment.
- The best fix is a concrete fact already present in the text: a number, a condition, a named source, a plain "est". The second-best is deletion. Inventing the fact is never a fix -- see the hard rules in Step 3.
- Targeted fixes only. Never rewrite the whole file, never rewrite a passage that carries no flagged tell.
- Examples in this file show the operation, not a target voice. Never carry their wording into the content.
- Adapt the blacklist check to the file's language. Each language has its own list above; for languages not listed (Italian, Spanish...), translate the patterns -- the tells are structural before being lexical.

Pour un contenu sensible, demandez d’abord un rapport et validez les corrections avec le texte source.

Transformer le prompt en skill pour Claude Code ou Codex

Pour une relecture récurrente, créez .../ai-pattern-review/SKILL.md :

---
name: ai-pattern-review
description: Strip AI writing tells from a content file — blacklisted vocabulary, inflated significance, structural and formatting patterns. Rewrites in place by default, or reports findings with --report.
disable-model-invocation: true
---

# ai-pattern-review skill

You clean AI writing patterns out of content files. Your job is to **read the content carefully**, find every AI tell, and fix it in place.

Sources: Wikipedia:Signs_of_AI_writing (en/de), Aide:Identifier l'usage d'une IA generative (fr), Wikipedia:AI_or_not_quiz, WikiProject AI Cleanup guide.

## Input

The user provides:

- **file** (required): path to the Markdown content file (e.g., `content/altimood.com/bivouac-montagne.md`)
- **mode** (optional): `--report` to list findings without touching the file. Default is to rewrite in place.

If no file is provided, ask the user which file to work on.

---

## Step 1: Read the file

Read the entire file content. Note the language (French, English, German, Italian, Spanish, etc.) to adapt your review accordingly.

---

## Step 2: Scan for AI patterns

...

La documentation Codex présente aussi $skill-creator. Avec les deux agents, reliez leurs dossiers au même fichier, comme dans mon organisation commune à Claude Code et Codex.

Commencez par le mode rapport sur vos propres textes. Les faux positifs montrent quelles règles adapter à votre voix.

Consultation stratégique et technique

Une chaîne éditoriale fidèle à votre voix

Je vous aide à documenter votre style et à vérifier les textes avant publication.

Email • RSS • Twitter

Être notifié des nouvelles publications

Poursuivre votre lecture