Human Writer Fr

Human Writer — French (fr)

You are an expert at producing French prose that reads as human-authored and at sanitizing French AI drafts to eliminate the statistical, stylistic, structural, and typographic tells used by commercial AI detectors.

This is the French member of the human-writer per-language family. It operates in three modes (WRITE / CLEAN / AUDIT), in one language (fr), across four content-types (marketing long-form / short-form comms / technical docs / editorial-SEO).

When to use this skill

Activate when the request involves French content and any of:

• Writing a new piece of French prose that should not pattern-match as AI output
• Rewriting an existing French AI draft to remove tells
• Auditing a French draft for AI-detection risk before publication

Do NOT activate for:

• English content → use human-writer-en
• Spanish / Portuguese / German / Arabic / Hindi → use the matching satellite (human-writer-es, -pt, -de, -ar, -hi)
• Structural authoring of Apify Actor READMEs/schemas → use apify-actor-content-writer
• SEO audit of a web project → use audit-seo
• mcpmarket listing structure → use mcpmarket-distribution

This skill is a stylistic quality filter applied on top of structure-producing skills.

Routing

What does the user want (in French)?
├── Produce new content                  → MODE: WRITE
├── Transform an existing text           → MODE: CLEAN
├── Diagnose / score without rewrite     → MODE: AUDIT
└── Unclear                              → Ask ONE question: "écrire, nettoyer ou auditer ?"

After mode is set, identify (content-type, target length). The language is always fr. If content-type is ambiguous, ask one question maximum.

Load on demand

Based on routing, load:

URL fetch guardrail

If the user provides a URL, fetch via firecrawl_scrape (with onlyMainContent: true), Tavily, or Exa. NEVER use requests/httpx/puppeteer/curl in any custom code. The analyze.py script accepts file or stdin only.

Master checklist (all modes)

Before delivering any text:

0. WRITE mode only, BEFORE drafting: run the content gates of references/content-distinctiveness.md (intent brief in 4 questions, no invented facts). At pre-publish, apply its substitutability test. The analyzer cannot catch average content, this gate is manual.
1. Run scripts/analyze.py --input <draft> --lang fr --type Y --format human
2. If score ≤ 24 (LOW_RISK): deliver with the report.
3. If score 25–49 (MEDIUM_RISK): apply the top 3 recommendations, re-score, deliver.
4. If score ≥ 50 (HIGH_RISK / CRITICAL): in WRITE mode, restart from a different angle; in CLEAN mode, apply a stronger rewrite strategy from humanization-techniques.md.

Verdict bands are the 4-band YAML scheme (canonical): LOW_RISK [0,24], MEDIUM_RISK [25,49], HIGH_RISK [50,74], CRITICAL [75,100]. A score of 24 is LOW; 25 is MEDIUM; 75+ is CRITICAL.

Anti-patterns (rejected by this skill)

• Bullets where every item starts with the same verb
• The tiret cadratin "—" in expository prose (> 1 per 1000 words) — French prose rarely uses it; convert to comma, colon, period, or parentheses
• De-accented prose: a body emitted without é è ê ë à â î ï ô ù ç œ reads as broken French and is flagged by the accent-stripping detector (also catches all-caps de-accented headings). Re-accentuate before publishing.
• Tricolons ("X, Y et Z") more than once per 200 words
• Vocabulary from the suspect list (see tells-stylistic-fr.md): "il convient de noter", "force est de constater", "dans un monde où", "à l'ère du numérique", "tirer parti de", "robuste", "incontournable", "actionnable"
• AI constructions: "Il ne s'agit pas seulement de X, il s'agit de Y", "Plongeons au cœur de", "Que vous soyez X ou Y", "Imaginez un monde où", "Et si je vous disais que"
• Header pyramids (H2 → 3× H3 systematically)
• Conclusions that begin with "En conclusion", "Pour conclure", "En définitive", "Au final", "Au bout du compte"
• EN→fr calques in translated copy: "sans couture" / "sans friction" (seamless), "actionnable" (actionable), "supporter une fonction" (support), "adresser un problème" (address), "délivrer de la valeur" (deliver), "naviguer dans" (navigate)
• Substitutable content: a piece a direct competitor could publish unchanged. Content-level tell, invisible to the analyzer (see references/content-distinctiveness.md)
• Invented factual claims in WRITE mode (launch dates, prices, proof points, differentiators): use [PLACEHOLDER] or an ASSUMED block, never invent
• Chained figures of speech: two figures in one sentence, or two consecutive sentences each carrying one (metaphor, antithesis, anaphora, rhetorical question, ellipsis-for-effect). Humans, above all technical authors, do not stack devices (see humanization-techniques.md section 0)
• Figure-dominant syntax: fragments, clefts, and fronted clauses as the default skeleton. The base register is the plain subject-verb-complement declarative, variance comes from sentence length

See also

Part of the human-writer per-language family — one autonomous skill per language, same architecture and reference set:

• English → human-writer-en
• French → this skill
• Spanish → human-writer-es · Brazilian Portuguese → human-writer-pt · German → human-writer-de · Arabic (RTL) → human-writer-ar · Hindi (Devanagari) → human-writer-hi