humaniseur-fr

Humaniseur : supprimer les patterns d'écriture IA du français

Your task

When given French text to humanize:

1. Clarify the target register first - Abbreviations, argot, orality, emojis and typography all depend on the expected niveau de langage. If it is not clear from the request or the input (soutenu, courant, familier ? for which medium ?), ask the user before starting to humanize
2. Identify AI patterns - Scan for all 38 patterns listed below
3. Rewrite problematic sections - Replace AI-isms with natural French alternatives
4. Preserve meaning - Keep the core message intact
5. Maintain voice - Match the intended tone and register
6. Add soul - Don't just remove bad patterns; inject actual personality (see Part 4)
7. Do a final anti-AI pass - Ask: "Qu'est-ce qui rend ce texte évidemment généré par IA ?" Answer briefly with remaining tells, then ask "Maintenant, fais en sorte qu'il ne le soit plus" and revise accordingly. Apply the three-signal rule: one isolated marker is noise (most are normal French), but three or more co-occurring in the same passage make a reader flinch — revise until no paragraph accumulates three

The 80 % rule: imperfect compliance is the point

Follow the humanizer rules most of the time — not always. It is fine to leave roughly 20 % of this skill's instructions unapplied.

• Real human prose contains flagged patterns at low density: « pour conclure » does not make a text 100 % AI, and a text that dodges every single tell with mechanical rigor is uniform in a new, equally detectable way.
• Transgress deliberately: keep an em dash that earns its place, one « par ailleurs », one tidy list.
• The tells are density and co-occurrence (see the three-signal rule), never a single occurrence.

One pass only. Do not run this skill repeatedly on the same text: each pass removes variance and injects its own habits, and quality degrades fast — by the second or third pass the voice the first one created is flattened again. If the result still smells AI after the final anti-AI pass, the fix is adding anchored content (a fact, a date, an opinion — see the limits note in Part 4), not another scrub.