Writing General

Writing register — the base

This skill is the base every writing task loads. It carries the shared rules that hold in every domain plus the general register for professional prose with no house style of its own.

Two domain skills sit on top of it and carry only what is additional:

• writing-legal — T14 law review articles, student notes, legal scholarship with footnotes.
• writing-econ — finance and accounting journal prose, working papers, job-market papers.

Load exactly one of those alongside this base when your prompt names that domain, and never import a rule across the domain line — the registers are contrastive by construction, and crossing that line is the single most damaging thing you can do here.

General register (general)

You are drafting and revising serious professional prose that is neither a law review article nor a journal submission: an SEC comment letter, a policy memo, a white paper, a letter to a regulator, a board memorandum. It has no house register of its own, so it borrows discipline from both scholarly registers without adopting either one's markers.

Everything below is measured, not asserted. The sources are two control corpora — 6,563 pre-2020 articles / 5,560,816 sentences from all 14 T14 flagship law reviews plus four business-law journals (/data/eh2889/aitic_corpus_law on rjds), and 11,198 pre-2017 articles / 8,733,332 sentences of finance and accounting scholarship (/data/eh2889/aitic_corpus). Percentages are share of sentences containing the feature. Rates written n/M are hits per million sentences.

Register: borrow from both, commit to neither

What the corpora say is NOT a register marker

Do not "fix" these. They are statistically indistinguishable between the two registers, so any advice keyed on them is style preference wearing empirical clothes.

• Passive voice: 7.91% law vs 8.55% finance. Both registers use it steadily and deliberately. Rewrite a passive when the agent matters, not on principle. Strunk's active-voice rule survives as a question ("who did this?"), not as a rule.
• however, 1.08% vs 1.01%; thus 0.75% vs 0.83%; moreover, 0.20% vs 0.16%. Connectives are not a tell in either direction.
• Em dashes: 0.51% vs 0.46%. Near-identical. The em-dash budget in prose-audit.py targets clustering inside a paragraph, which is a different claim from the overall rate.

Two measurements that look tempting and are confounded — ignore them:

• Sentence length (mean 10.3 law / 10.9 finance): an artifact. The corpora are fitz.get_text() PDF output, which breaks lines mid-sentence, so measured length is a floor, not a distribution.
• Contractions (4.04% vs 0.86%): the regex catches possessive 's, and legal prose is dense with the court's, plaintiff's. It is not measuring contractions.

Conventions

• Lead with the ask. A comment letter's first paragraph says what you want changed and why. The analysis follows; it does not build up to the point.
• Number the sections and refer to them by number.
• Cite in full the first time, then by short name. No supra, no author-date, unless the document is going to a venue that expects one.
• One word per concept. No synonym cycling — if it is the "passive block" in Section 2, it is not the "index cohort" in Section 5.
• Attribute every number. A quantity with no source is not evidence.

The shared base — applies to all three registers

The base layer: Strunk, run through the corpora

The source guide is Strunk & White, The Elements of Style, vendored in full at ${CLAUDE_PLUGIN_ROOT}/skills/writing/references/elements-of-style.md — read it there for the full text of any rule below. Where this file and that guide disagree, this file controls: its prescriptions were checked against all 14,294,148 sentences of the combined corpora and split three ways. A prescription that human scholars violate constantly is not thereby wrong — these guides are prescriptive — but a rule in the second bucket fires on roughly one sentence in fifteen, and a preloaded rule that noisy is worse than no rule.

Ship — cost-free, so treat these as rules

The measured rate is how often the phrase appears in the combined corpora. A rate under ~50/M means enforcing the rule costs essentially nothing, because almost nobody writes it. — means the swap was not measured: it is a low-risk judgment call, not a finding.

Advisory — real signal, constant in human prose, so judge in context

Every one of these is idiomatic at a rate that makes a hard rule pure noise. Flag them when a sentence is genuinely worse for them; never rewrite on sight, and never report a run of them as a finding.

Dropped — these were register mistakes

Each of these guide rules, applied to its own domain's prose, damages the draft. They are recorded here so nobody re-derives them from the source guides.

Prohibited constructions — owned by ai-anti-patterns

The tic dictionary lives in the ai-anti-patterns skill, which every prose agent loads alongside this one; prose-audit.py flags each entry with a span id, so cite the span rather than re-scanning by eye. /ai-tic <phrase> is how a candidate becomes a rule — it runs the FP-hunt against both corpus halves and refuses anything over that threshold. Do not restate tics here: a second copy cannot be regenerated from the dictionary and goes stale silently.

Phrases the corpus VINDICATED — use them freely

These read as AI to many readers and are in fact standard scholarship. Do not let a reviewer talk you out of them, and do not "fix" them in someone else's draft.

Formatting

• Prose, not bullets. For reports, documents, technical documentation, and explanations, write prose without bullets, numbered lists, or excessive bolding, unless the person asks for a list or ranking. Use lists, bullets and formatting only when (a) asked, or (b) the content is multifaceted enough that they are essential for clarity.
• No bold inline headers opening a paragraph (**The objection.** Text follows…, #strong[…], \textbf{…}). Use a prose topic sentence, an italic label, or a real heading. List items are exempt by design, and so is bold marking a genuine defined term.
• No bold on bare numbers. Emphasize the claim, not the digits. This is the densest formatting tell measured in a real draft: 32 of 66 bold spans in one comment letter were bare quantities.
• No emojis. Ever, in a draft. (A slide deck is not a draft.)
• No ALL-CAPS for emphasis on ordinary words (is NOT a separate cut). Acronyms and table headers are fine.
• Do not hard-wrap prose for a soft-wrapping reader. The test is who reads the text, not the file extension.
  • Soft-wrapping reader — an email body, an Obsidian note, a web form, a chat message, anything rendered rather than read as source: one paragraph, one line, no manual breaks at any column. Those readers reflow to the pane, so breaks at 80 (or any) columns land mid-sentence at whatever width the reader uses, and a one-word edit turns into a re-wrap of the whole paragraph. Let the editor wrap it.
  • Fixed-width reader — a commit message (72 columns by convention), a code comment, a SKILL.md, a .typ or .tex source file: wrapping is correct. Keep it. This file is wrapped at ~100 columns for exactly that reason, and prose-audit.py exempts such sources by name, suffix and frontmatter shape rather than flagging its own register.

Before you call a draft done

Run the deterministic audit and cite span ids rather than re-reading by eye:

uv run --with lxml --with pyyaml python3 ~/projects/workflows/scripts/prose-audit.py \
  --json --style legal|econ|general <draft>

hard spans block; soft spans are advisory. To test a phrase you suspect is a tic, use /ai-tic <phrase> — it hunts both corpus halves and will tell you when your instinct is wrong, which is most of the time.