Humaniseur FR

The French half of humanizer. English text is not this skill's business: use humanizer for that, and translation when the job is moving a string between the two languages.

Register is the first rule

Never lower the register of the input. An input in langage soutenu stays soutenu. Formal is not the same as machine-made: a subordinate clause, a precise connector and a long periodic sentence are features of good French, not artefacts. Only what is genuinely mechanical comes out — inflated significance, copula avoidance, synonym cycling, promotional filler.

So step zero is the target register (soutenu, courant, familier) and the medium. When the request does not say, ask before rewriting anything.

The dose, and the number of passes

• The 80% rule. Applying every rule all the time is its own tell. Keep an em dash that earns its place, one par ailleurs, one tidy list.
• The three-signal rule. One marker is noise — most of them are ordinary French. Three or more in the same passage is what makes a reader flinch. Judge density and co-occurrence, never a single occurrence.
• One pass. Do not run the skill twice over the same text; the second pass flattens the voice the first one created. If it still smells machine-made after the final pass, the fix is anchored content — a fact, a date, an opinion — not another scrub.

It yields to a more specific method

A LinkedIn post, a press release, ad copy, developer documentation in a README: whatever rules those bring win on conflict, and this skill covers only what they leave open. A short punchy paragraph, a tidy list or a heading is the format speaking, not the machine.

French is not English with different words

• The AI lexicon is its own, and crucial is its first word, not "delve".
• Anglicisms are a major tell, because the model thought in English first.
• Typography is strict, and it is checked here rather than left to a style guide: spacing, guillemets, number format.
• The dissertation tradition (thèse / antithèse / synthèse) overlaps with AI structure, so announce-and-balance is school-legitimate and passes unnoticed.
• Longer sentences are normal. "Vary the sentence length" therefore means something narrower: the missing extremes are the tell, not the mean.

The patterns

1 · Content

1 · Inflation of significance. constitue un tournant, témoigne de, joue un rôle crucial, souligne l'importance, ouvre la voie à, un paysage en mutation, une empreinte indélébile. State the fact; drop the significance.

2 · Notability and coverage. cité par, a list of prestige outlets, une forte présence sur les réseaux sociaux. Use the real source and what it said, or cut the list.

3 · Superficial -ant participles. soulignant, mettant en lumière, reflétant, contribuant à, favorisant, englobant. The French shape of the English -ing rider.

4 · Promotional language. vibrant, riche, profond, niché, au cœur de, révolutionnaire, incontournable, un joyau. Also the valorising adjective or inclusive doublet absent from the source — nos soldats becoming nos vaillants soldats: restore the source's wording.

5 · Vague attribution. les experts estiment, les observateurs soulignent, il est largement reconnu que, des rapports sectoriels. Name the source, or drop the claim.

6 · The défis et perspectives sandwich. Malgré …, fait face à, En dépit de ces défis, Perspectives d'avenir, L'avenir s'annonce prometteur. End on a concrete fact instead.

2 · Language, grammar and style

7 · Overused AI vocabulary. The one word list, and the only place a word is a tell by itself.

AI	strategy
crucial, essentiel	a domain term, or drop it
également (the top measured marker)	aussi, de même, or drop — one a paragraph at most
défi	problème, difficulté, or name the obstacle
significatif, robuste, substantiel	a number
holistique	delete
compréhensif (for exhaustif)	exhaustif, complet
disruptif	de rupture, or describe the change
notamment (over one per 800 words)	en particulier, entre autres, or restructure
par ailleurs, en outre, de plus	or, reste que, n'empêche que
il convient de noter que	delete, start the sentence
dans le paysage …, au cœur de	delete, or name the place
la pierre angulaire, un levier puissant	say what it is
permettre de, favoriser, optimiser	the concrete verb
mettre en lumière	montrer, révéler
naviguer dans, déverrouiller le potentiel de	describe the action
garantir, assurer, offrir	state the fact
dans cette optique, à cet égard	delete, or link concretely
que vous soyez X ou Y	address the actual reader

Add to it: ministerial jargon outside an administrative text (dispositif, acteurs, enjeux, mise en œuvre, dynamique territoriale); formulaic openings (Dans le paysage …, À l'ère de …, Dans un monde où …, Il est essentiel de noter que …, Plongeons dans …, Découvrez comment …); formulaic closings (En conclusion, En résumé, En somme, Au final); and the connectors the model almost never uses, which are therefore free authenticity: Or, Quoi qu'il en soit, Toujours est-il que, Force est de constater que, Reste que, N'empêche que, Soit dit en passant.

8 · Copula avoidance. constitue, fait office de, se positionne comme, dispose de, offre. Use est and a.

9 · Negative parallelism. Non seulement … mais aussi …, Il ne s'agit pas seulement de … mais de …, Ce n'est pas un simple X, c'est un Y.

10 · Systematic rule of three. Ideas forced into groups of three.

11 · Synonym cycling. One referent under three names — le protagoniste, le personnage principal, la figure centrale.

12 · False scales. de X à Y, de A à B where neither pair is a scale.

13 · Anglicisms of architecture. The most reliable family, because the model reached for English first.

AI	French
faire du sens	avoir du sens
adresser un problème	traiter, aborder
implémenter (outside IT)	mettre en œuvre
impacter	affecter, toucher
supporter (to support)	prendre en charge
définitivement (for "certainly")	sans aucun doute
basiquement	en gros, fondamentalement
a comma before et	no comma before et

14 · Redundant adjective doublets. crucial et essentiel, robuste et fiable, innovant et avant-gardiste.

15 · Em dashes. French prefers the comma and the parenthesis for an incidental clause. Weakened tell: readers know it now and humans use it, so reduce the overuse without treating a sighting as proof.

16 · Mechanical bold, headed lists and tables. - **Titre :** … bullets, and a table presenting ordinary prose as data. Convert to prose; keep a table when the content is genuinely tabular.

17 · English title case in headings. French capitalises the first word only — ## Négociations stratégiques, not ## Négociations Stratégiques.

18 · Emoji. A register feature, not a defect. Keep the one that carries tone or replaces words; remove the one that decorates structure. The tell is the systematic series — same position, one per bullet, one per heading. Formal documents take none. When the author's voice is unclear, ask.

19 · Quote and mark inconsistency. The tell is not one variant but the mix: straight quotes, curly quotes and chevrons in one text, or ' and ’ alternating. Match the register — chevrons « … » with non-breaking spaces in worked official prose, curly quotes elsewhere — and keep it consistent.

20 · Conversation artefacts. J'espère que cela vous aide, Bien sûr !, Absolument !, Vous avez tout à fait raison, Souhaitez-vous que …, N'hésitez pas à, Voici un …. Remove the wrapper, keep the content.

21 · Knowledge-limit clauses. Selon les informations disponibles, Bien que les détails spécifiques soient limités, sur la base des données accessibles. State what the source does not show, or cut the sentence.

22 · Filler. Afin de parvenir à cet objectif → Pour y arriver; En raison du fait que → parce que; À l'heure actuelle → aujourd'hui; Dans l'éventualité où → si; Le système a la capacité de → le système peut; Il est important de noter que → delete and start the sentence; En ce qui concerne → sur, quant à.

23 · Excessive hedging. On pourrait potentiellement arguer que cette politique pourrait éventuellement avoir un certain effet.

24 · Generic positive conclusions. L'avenir s'annonce prometteur, poursuit son chemin vers l'excellence, un pas majeur dans la bonne direction.

25 · Structural uniformity. Paragraphs of near-identical length, lists of 3/5/7/10 items, an invariable intro-body-conclusion. The only quantified French study found mean sentence length nearly identical — 21.0 words human against 21.7 AI — and AI producing almost no sentence under 15 or over 39 words. Reintroduce the tail. Also here: sentence-start anaphora (Cela …, Cette approche …, Ce système … one after another), and the LinkedIn register — one-sentence paragraphs, Et là… tout a changé, relentlessly upbeat — which now reads as machine-made even when a human wrote it.

26 · Residual Markdown and technical artefacts. Unrendered **mot** or ## in a medium that does not render Markdown, - **Titre :** bullets pasted the same way, :contentReference[oaicite:2]{index=2}, a leftover refusal (Je suis désolé, mais je ne peux pas …). These are the strongest tells of all, because no spell-checker produces them — only a pasted chat output does. Strip zero-width characters (U+200B, U+200C, U+200D, U+FEFF). Never strip non-breaking spaces (U+00A0, U+202F): they are correct French typography before ; : ! ?, and confusing the two raises a false alarm on every properly typeset French text. Markdown in a README, a wiki or developer documentation is the medium, not a tell.

3 · Discourse architecture

Lexical scrubbing is not enough. A text with zero flagged words can still read machine-made because of how it is built. These are craft heuristics rather than measured figures.

27 · Announcement, recap and echo of the brief. An intro announcing what the text will say, a conclusion repeating it, headings mirroring the announced plan one for one, a first sentence restating the question asked. Start in medias res; end where the intro could not have predicted.

28 · Catalogue structure. Sections that could be reordered without damage, every aspect of the topic at equal depth. The permutation test: if two sections swap without damage, it is a disguised list, not an argument. Choose an angle, cut the aspects that do not serve it, make the order necessary.

29 · The paragraph mould. Topic sentence, two or three supports, a mini-conclusion, in every paragraph, with D'abord / Ensuite / De plus / Enfin scaffolding. Drop the connectors; let one idea straddle a paragraph break.

30 · False balance. Every claim immediately counterbalanced — Cependant, il convient de nuancer …, a symmetrical d'une part / d'autre part, a tout dépend du contexte conclusion. Take a position and nuance once. Deletion test: remove the objection; if the conclusion still holds, it was ornament. Quebec caveat: rédaction épicène and OQLF plain-language norms push human institutional writers toward exactly that flat, symmetrical shape, so balance alone is not evidence there.

31 · SEO over-sectioning and the ghost Q&A. A heading every two paragraphs, a table of contents on a short text, an FAQ block, Pourquoi est-ce important ? Parce que …. A heading governs four or five paragraphs or it merges; a self-question becomes a statement.

32 · Constant granularity. The whole text at one altitude — no date, no name, no price, no error message. Force one verifiable, named, dated detail per section. An author with none has a content problem, not a style problem.

33 · No holes. Every question the text raises gets answered, no thread left open. Real expertise leaves holes, because the author knows where the knowledge stops. Leave one question honestly open and name the limit.

34 · The list as avoidance. A bullet list appearing exactly where the reasoning gets hard — the decision, the trade-off, the prioritisation. Ask what decision the list avoids, and write the sentence that chooses.

35 · No occasion. Nothing explains why the text exists now, for whom, or triggered by what. Anchor it in its occasion within the first paragraphs. When there is none, ask the author for it.

4 · Soul

Removing tells is half the job. Sterile text is as suspicious as crucial.

Know the limit. Experienced French readers no longer rely on style. What they trust is behavioural: publishing cadence, whether the author exists, whether the sources check out. No amount of scrubbing fixes a text with nothing situated in it. When the input has no anchored content, ask for one real example, one date, one source, and build the rewrite around them rather than polishing the surface.

Then add voice: an opinion (Franchement, je ne sais pas quoi en penser beats a neutral pros-and-cons); unequal rhythm, short sentences against long periodic ones; je where it fits; a parenthesis or a tangent; irony, understatement and self-deprecation; one well-placed rare word, soutenu or argotique, matching the register; the regionalisms of the author's actual French, never a forced one; in familier only, the dropped ne and the contraction; and the abbreviation a human reaches for (pb, tjs, 14h30, A+, cf., PS:, RDV). The model never writes je ne sais pas, so a frank admission of not knowing is among the strongest authenticity markers — as is humour. Both belong only where the context allows: a legal notice tolerates neither.

The test for "soulless even when clean": every sentence the same length, no opinion, no uncertainty, no first person where it would fit, no humour, reads like a press release.

Procedure and output

1. Establish the register and the medium. Ask when the request does not say.
2. Read the input once and mark the patterns, strongest first.
3. Rewrite, then add voice.
4. Self-audit: Qu'est-ce qui rend ce texte évidemment généré par IA ? Answer in two or three lines, then fix what the answer names.
5. Return, in this order: the final version (house order is conclusion first), the two or three tells that remain, then the patterns removed. The draft is available on request.

Never add a fact. A date, a number, a quote or a name the source does not carry is a hallucination, not a human touch.

In this fleet

• French has two homes: a .po catalogue target and a reply to the owner. French anywhere else in the repository is drift the next reader pays for — the translation skill's rule, which this skill does not relax.
• Never reword an existing i18n string. The catalogue is regenerated from the source string.
• Do not translate what WordPress translates. article, page, extension, tableau de bord: where WordPress already renders the term, a second name for one concept is a defect.
• Check the typography on the way out: non-breaking space before : ; ! ?, and no title case in headings. translation states both; this skill enforces them after a rewrite, which is the moment they get broken.

Verification

The skill was followed correctly when:

• the register of the output is the register of the input;
• it reads aloud as French a French speaker would write;
• no fact, date, number or source was invented;
• typography is consistent and correct — spacing, guillemets, number format;
• the sentences have a short tail and a long tail, not just a mean;
• the result came out in one pass.

What this skill kept from upstream

• Register first, and never lowered. Upstream's strongest rule, and the one almost every other humanizer breaks by turning prose casual.
• The three-signal rule and the 80% rule. Density and co-occurrence decide, never a single occurrence — which is what stops the skill producing sterile prose.
• One pass only, with the reason: a second pass erases the voice the first created.
• The pattern taxonomy, compacted from 38 entries, with the French triggers kept verbatim because the triggers are the tool.
• Part 4, soul. Upstream is unusual in saying that removing tells is only half the job, and in naming the behavioural signals no style fix reaches.
• The typography details: zero-width characters out, non-breaking spaces in.
• The measured French figures — également at four times human rate, the sentence-length distribution, the tense and pronoun profile — because they are the only quantified evidence in the source.

What was dropped, and why

• The French prose itself. Upstream is written in French; everything written to this repository is US English, so the reasoning is in English and only the terms, triggers and examples stay French. translation is written the same way.
• The full worked example. A thirteen-line avant and three versions of it, about AI coding assistants, then a summary of the twelve patterns removed. The patterns are stated in the list; the example was illustration and long.
• The per-region slang and abbreviation palettes in full. France, Belgium, Quebec, Suisse romande — about forty items across four lists. Kept as a rule with a handful of examples, because a palette is only usable when it matches the author's real French, and a wrong regionalism is worse than none.
• The narrative-fiction register (pattern 29: lyrisme de pacotille). This fleet writes documents, translations and replies, not French fiction.
• allowed-tools, license, compatibility, metadata, the OpenClaw block and the freshness warning. Upstream's harness fields, meaningless to Reasonix; the freshness point survives in the weakened-tell notes.
• The six measured percentages for devoir / continuer / tenir / ensemble. Six figures for one tendency, kept as the tendency.

What upstream was missing

• The house homes for French. A .po target and a reply to the owner, and the translation skill's rule that French anywhere else is drift.
• The i18n rule. Rewording an English source string invalidates every translation of it, and no upstream prose skill knows a catalogue exists.
• WordPress vocabulary. article, page, extension, tableau de bord — terms WordPress already translates, so translating them again produces a page where one concept has two names.
• The output order. Upstream returns draft, tells, final, summary. House order is conclusion first: the final version, then the tells that remain, then what was removed.