copywriting-prose-creator

Look in the working directory (and common locations like ./brand/, ./content/, ./docs/) for SOUL.md, TONE.md, prior PROSE.md, and any content corpus. If SOUL.md or TONE.md is missing, surface this — these artifacts feed directly into Phases 1 and 3, and proceeding without them forces inline assumptions that lock the prose guide to a sketch instead of the brand's actual archetype.

If missing, offer two paths:

1. Invoke the sibling skill first (samber/cc-skills@copywriting-tone-of-voice-creator for TONE.md). Why: TONE.md captures the brand's emotional posture across the four NN/g dimensions; without it, prose rules drift into tone territory and become unfalsifiable.
2. Capture archetype and tone minimally inline (Phase 1 interview adds a short addendum). Pragmatic for one-off prose audits.

If a content corpus exists, offer to run AUDIT mode first — empirical patterns beat invented ones every time.

Ask in 2–3 batches. Skip any field already supplied by SOUL.md, TONE.md, or prior conversation context. Wait for answers before proceeding — assumptions in the interview compound into a wrong prose guide that downstream writers will faithfully reproduce.

Required fields (full battery in references/discovery-questions.md):

• Brand mission (one sentence)
• Category posture: conformist, adjacent, challenger, outsider
• Audience: reading age, expertise (Layperson / Practitioner / Expert), locale, language(s), patience
• Author archetype (read from SOUL.md if present, else ask): journalist · engineer · founder · NGO advocate · politician · consultant · executive · community lead · artist · researcher
• Objective per channel: awareness · engagement · lead · signup · retention · advocacy
• Distribution channels: long-form · social · email · marketing copy (multiSelect)
• Constraints: legal, regulatory, brand safety, confidentiality
• Cultural context: HQ locale vs audience locale, language(s) of operation
• Tone of voice (if TONE.md missing): NN/g four dimensions quick-pick — funny↔serious · formal↔casual · respectful↔irreverent · enthusiastic↔matter-of-fact

Match the brand to one of the 11 covered categories. Load the playbook from references/category-playbooks.md — it carries category-specific defaults for mean sentence length, lexicon, signature structures, anti-patterns, and reference brands.

Uncovered context → delegate research. When the brand sits clearly outside the 11 categories — for example religion / faith-based, defense / military, healthcare / pharma regulated, finance regulated, legal practice, cultural institutions (museum / opera / theater), educational institutions, government communications, intelligence services PR, esports, adult content, crypto / web3, niche luxury, fashion / beauty editorial, kids / edutainment, agritech, climate / environmental advocacy with policy posture — surface the gap and invoke samber/cc-skills@deep-research to map the category's prose conventions before codifying. Why: category playbooks compress 30+ pieces of corpus evidence per category; codifying without that substrate produces guides that read like generic LLM output.

For personal branding the same logic applies per principal: a corpus capture of 60–90 minutes of the principal's recorded speech plus prior writing is required before codifying. Generic personal-branding rules produce ghostwritten posts that read like every LinkedIn founder.

Codify each layer in order. Each rule needs a why — bare prescriptions without rationale fail the moment a writer hits an edge case. Detail rules and examples in references/five-layers.md.

1. Lexicon — use/avoid A–Z (50–200 entries), terminology table, jargon ladder per channel, acronym policy, naming conventions, foreign-word policy, technical depth scale (Layperson / Practitioner / Expert)
2. Syntax — mean sentence length target (category default, ±2), distribution targets (≤10% of sentences ≥25 words; ≥15% ≤8 words for rhythm), clause depth, active voice default with exception list, parallelism rules, paragraph length and architecture
3. Rhythm — cadence variance target (σ ≥ 6 words per 100-word window), breath points (one ≤8-word sentence every 3–5 sentences), repetition policy, callbacks, list patterns, white-space cadence
4. Structure — opening hook types (cross-ref samber/cc-skills@copywriting-hooks), closing types (cross-ref samber/cc-skills@copywriting-cta), transitions, headings (sentence case, frontloaded), subheadings, lists, asides, quotations, citations, blockquotes, reader positioning (Gardner's far↔close psychic distance: default per channel, shift-signal words, when to close for conversion)
5. Voice markers — 5–12 signature moves, signoffs, recurring metaphors, idioms, taboos, intentional tics (all rationed; unrationed markers collapse into self-parody)

Diagnose the corpus before locking the targets:

1. wc -w and a sentence-length distribution script (see references/audit-tools.md) — establish current mean and σ before declaring targets
2. Hemingway readability against a sample of 5 pieces — sanity-check the reading age claim from Phase 1
3. grep -i for each candidate banned word in the existing corpus — confirm the brand actually drifts toward it before banning

Two non-negotiable tables.

Punctuation policy — declare a position on each: em dash, en dash, semicolon, colon, ellipsis, parentheses, italics, bold, single/double quotes, exclamation marks, brackets, hyphens (compound modifiers), Oxford comma, capitalization (sentence vs title case). Defaults and rationing tables live in references/five-layers.md.

Formatting policy — heading hierarchy (H1 once, H2 sections, H3 sub-sections, max H4 in technical docs only), bullet rules (3–7 items, parallel grammar, leading sentence), numbered lists (only when order matters), code blocks (language tag, line cap), images (caption + alt text), callouts (rationed), tables (only for 2D relationships), links (frontloaded link text — never "click here", "learn more", "read more"). Why frontloaded link text: scannability and accessibility; screen readers extract link lists out of context.

For each in-scope channel grouping (see table above), produce a CHANNEL section in PROSE.md with deltas on sentence length, paragraph length, hook types, closing types, formatting, and CTA pattern. Pull the transformation rules from references/channel-adaptation.md.

Generic groupings keep PROSE.md portable: when a brand adds a new platform within a grouping (e.g. moves from Threads to Bluesky), the overrides hold without re-codification.

• English variant: declare US / UK / international English (spelling, punctuation, date format)
• French ↔ English: list the few French words permitted in English text (raison d'être, savoir-faire) and forbid others without translation; conversely declare English loan-words accepted in French (le marketing, le briefing) vs taboo
• False cognates: éventuellement ≠ eventually, actuellement ≠ actually, important often ≠ important; full list in references/multilingual.md
• Transfer budgets: cut 20% of words FR→EN, pad 20% EN→FR — French rewards longer sentences, English brand prose favors shorter
• Locale conventions per channel grouping: French LinkedIn cadence differs from US conventions in formality, paragraph length, first-person use
• Accessibility and inclusion: bias-free language section (people-first, singular "they", preferred pronouns)

For multilingual brands: one PROSE.md per language, not a translated single guide. Maintain a mapping document of shared pillars and divergent rules.

The dominant prose-drift risk in content factories is convergence on LLM-default register. Codify rules LLMs do not follow by default — that is the durable defense.

Full inventory in references/anti-patterns.md. Headline patterns:

• Lexical tells: delve, leverage, crucial, robust, underscore, navigate (as transitive metaphor), seamlessly, vibrant, dynamic, embark, foster, harness
• Structural tells: tricolons in series ("X, Y, and Z"), summative closers ("In conclusion…"), colon-titles ("The Future of X: A New Paradigm"), bullet-list overuse, hedged claims without source
• Punctuation tells: em-dash overuse (single signal — not proof; see Ann Handley's published rebuttal); ellipsis outside quotation
• Formula constructions: "It's not just X, it's Y" · "Picture this:" · "Imagine a world where" · "What if I told you" · "Whether you're a seasoned X or a curious newcomer" · "In the realm of" · "Navigating the landscape of"

Diagnose LLM drift quantitatively:

1. grep -c -iE 'delve|leverage|crucial|robust|underscore' across the corpus — frequency ≥1 per 500 words is a strong tell
2. Sentence-length σ < 4 across a 100-sentence window — uniformity is a stronger tell than any single lexical signal
3. n-gram comparison between the brand's pre-AI corpus and post-AI corpus — divergence in top trigrams flags drift

Detection is unreliable as a single source of truth. Use these as triage, not verdict. The Stanford HAI / Liang et al. (2023) work showed GPT detectors misclassify TOEFL essays by non-native English writers at headline rates above 60%. Treat any single signal as suspicion, not proof.

Use the hybrid template in references/prose-md-template.md:

1. Narrative sections for each layer + policy (the why and the how)
2. Do/don't tables as an annex (the quick-reference scan layer)
3. Sample bank: ≥10 before/after pairs, ≥3 exemplar pieces if provided, hook bank and closing bank cross-referenced from samber/cc-skills@copywriting-hooks / @copywriting-cta
4. Cross-references to TONE.md and SOUL.md (read together, not in isolation)
5. Versioning footer: semver, date, owner, changelog stub

---