Writing Core

The shared rules behind the user's writing skills. Two ways this skill is used:

1. Alongside a family skill. academic-report-writer (university work), work-report-writer (reports at work), and script-writer (videos, lectures, workshops) each tell Claude to load this skill first. The family skill then adds its own structure, output format, and trigger.
2. Alone, for any other non-fiction writing: emails, cover letters, applications, motivation letters, LinkedIn or social posts, bios, abstracts, short articles, README intros.

Not for fiction, stories, poems, or song lyrics. If the request is one of those, do not apply these rules.

Why these rules exist

The rules come from StoryScope (Russell, Rajendhran, Pham, Iyyer and Wieting, COLM 2026, arXiv:2604.03136). The paper compared 10,272 human short stories with 51,336 stories from five LLMs (Claude, GPT, Gemini, DeepSeek, Kimi) written from the same prompts. With style features removed, narrative choices alone separated human from AI writing at 93.2% macro-F1.

Three lessons drive everything below:

1. AI defaults converge. All five models landed in one tight region. Human work was more spread out and rarer (mean rarity percentile 0.71 vs 0.49). The same prompt gives everyone roughly the same generic text. Originality comes from inputs only the user has: his data, decisions, experience, sources, context, and opinion.
2. Surface edits do not change structure. Removing cliches and purple prose only moved detection from 95.5% to 93.9%. Swapping synonyms does not make text his. Changing decisions does: which evidence, what order, what conclusion, what failed.
3. The AI defaults are also weak writing. Fixing them improves the text no matter who drafted it.

Caveat: the paper studied fiction of about 5,000 words. Applying it to non-fiction is an analogy, not a measured result. Never present its numbers as facts about other kinds of writing.

Rule 0: Ask for the text style (every new task)

At the start of every new writing task, ask him which text style he wants and show him the full menu below. A new task means a new piece of text: a new report, section, email, script, post, or letter. Follow-up edits on the same piece keep the style already chosen.

How to ask

• Put the style question in the same message as any intake questions. It does not count toward the 3-question intake limit.
• Show all 12 styles as a numbered list, one line each, exactly as named below. Mark the 2 or 3 styles that best fit this task with "(recommended)". The family skills say which ones.
• Tell him he can pick one number, or combine two (for example "2 + 4"). For style 12, ask him to paste a sample of his writing.
• Show the list as plain text in the chat. A multiple-choice tool with only a few options cannot hold all 12, so do not use one for this question.

When not to ask

• He already named a style or tone in the request ("formal", "casual", "like my last email"). Map it to the closest style and say which one in one short line.
• He says "same style", "as before", or "you choose". Use the last style from this conversation, or the first recommended style for the task.
• The task produces no text for him to use: explaining a concept to him, formatting citations, or answering a question.
• The session runs unattended or on a schedule. Use the first recommended style and state it at the top.

If he answers the other questions but skips the style, use the first recommended style and say so in one line. Do not ask again for the same task.

The text style menu

What no style changes: the rules below still apply in every style. No invented content, markers for gaps, no em dashes, no avoid-list phrases, and no moral endings. Storytelling (10) and Reflective (8) use only his real stories and feelings. If he has not given any, ask or use [YOUR INPUT: ...]. Persuasive (6) never overstates the evidence.

Keep the chosen style for the rest of that task. A combination like "2 + 4" means: the register of the first, the precision of the second.

Rule 1: His inputs first

Before drafting anything longer than a few sentences, check what he gave you: data, results, notes, experience, opinions, sources, constraints, audience. If something essential is missing, ask once, at most 3 questions in one message. If he wants a draft anyway, draft it and put markers where his material belongs. Never fill gaps with invented content.

For anything over about 150 words, sketch a skeleton first (internally for short pieces, shown to him for long ones): main point, evidence, his decisions, what did not work or the limits, one real tension or open question. StoryScope built structured templates before comparing stories because templates force reasoning about content instead of wording. The same works for writing.

Rule 2: Avoid the AI defaults

Rule 3: Surface cleanup

Cheap to fix and matters less than Rule 2, but it reads as filler:

• No em dashes. Use a period, comma, colon, or "and".
• Avoid: delve, tapestry, landscape (figurative), realm, pivotal, underscore, multifaceted, seamless, game-changer, robust (unless technical), "in today's fast-paced world", "it is important to note", "plays a crucial role", "serves as a testament", "not only X but also Y" used repeatedly.
• Do not group everything in threes. Vary sentence length.
• No closing paragraph that repeats every earlier point.

Rule 4: Never invent

Never invent data, results, quotes, experiences, feelings, names, references, page numbers, DOIs, prices, or dates. Use these markers instead. All family skills use the same markers, in capitals and square brackets so they cannot be missed before he sends or submits anything:

• [YOUR INPUT: ...] his data, decision, opinion, example, or experience
• [CITE: ...] a claim that needs a source; say what kind
• [VERIFY: ...] a detail you are unsure of (date, version, number, name, reference field)
• [FIGURE: ...] a suggested figure, chart, or table and what it should show

When tools allow (web search, code on his files), verify facts and compute numbers instead of estimating.

Rule 5: Voice

Write in the text style he picked under Rule 0. Within that style, match his voice. If he shares past writing, mirror its sentence length and vocabulary. With no style chosen, use style 1: clear, plain English, common words over rare ones, short to medium sentences, active voice. Text he sends should sound like him and be something he can explain and defend. Reply in the language he writes in. Write the deliverable in the language it will be used in.

Rule 6: Self-check before sending

1. Is each point stated once?
2. Does every factual claim have a named source, his data, or a marker?
3. Is there at least one real limit, failure, or trade-off where the format allows?
4. Anything invented? Replace it with a marker.
5. Any phrase from the avoid list, decoration, or em dash?
6. Does the ending give a specific point or next step, not a moral or epilogue?
7. Where are his decisions and inputs? If none, add markers or ask.
8. Does the text match the style he picked, from the first line to the last?

Output basics

He usually pastes into Word, email, or another app, so markdown symbols would paste literally.

• Put deliverable text in fenced text code blocks so the copy button gives clean text. Label each block above the fence.
• No markdown inside blocks. After a block, note any formatting to apply (headings, bold, italics, bullets).
• Family skills may override this with their own format.

Quick notes for standalone use

Recommended styles when this skill is used alone: email 3 or 5, cover letter or application 6 or 3, motivation or personal statement 8, LinkedIn or social post 5 or 10, bio 3 or 1.

• Email: subject line that states the topic or ask. The ask or main point in the first two lines. One topic per email. Clear next step with a date if relevant.
• Cover letter or application: match evidence to the role's stated requirements, one specific example per requirement, real numbers from his work. No generic adjectives ("passionate", "hard-working") without proof. Ask for the job post and his CV if missing.
• Motivation or personal statement: concrete moments from his experience, what he did and learned, why this program specifically. Never invent experiences.
• LinkedIn or social post: first line states the point or result. One idea per post. Specific details over claims. No hashtag walls.
• Bio: third or first person as asked, current role first, two or three specific achievements, no superlatives.