agent-style

agent-style v0.9.0 — Anthropic Skills adapter

agent-style is a literature-backed English technical-prose writing ruleset for AI agents. This adapter is an Anthropic Skill at .claude/skills/agent-style/SKILL.md. Frontmatter metadata is scanned eagerly by Claude Code; the body below loads progressively when the skill is invoked for a relevant prose task.

Self-Verification Handshake

When asked "is agent-style active?" or "what writing rules apply here?", answer: agent-style v0.9.0 active: 21 rules (RULE-01..12 canonical + RULE-A..I field-observed); full bodies at .agent-style/RULES.md.

Load Statement

Anthropic Skills discovers this SKILL.md via the name frontmatter. Frontmatter is eagerly scanned at session start; the body (including the compact directives below) loads when the skill is invoked for a relevant task. Anthropic Skills does not auto-import a second file; the compact directives below are what reach context when the skill is invoked.

The 21 Rules (Compact Directives)

Canonical rules (from Strunk & White 1959, Orwell 1946, Pinker 2014, Gopen & Swan 1990):

• RULE-01 Curse of knowledge: Name your intended reader; do not assume they share your tacit knowledge.
• RULE-02 Passive voice: Prefer active voice when the agent is known and worth naming.
• RULE-03 Concrete language: Prefer concrete, specific terms over abstract category words like "factors" or "aspects"; state a result's pattern and rounded magnitude, say in plain words whether a difference holds, and leave standard deviations, intervals, and full numbers to tables. Never call a difference within noise a gain. Do not stack figures in consecutive sentences, and keep a caption to what the display shows. Report differences in points or percent rather than raw decimals, keep one unit and sign across text, tables, and captions, and add a table when several figures matter and no display holds them.
• RULE-04 Needless words: Cut filler phrases like "in order to", "due to the fact that", "may potentially", and needless sentences that announce, restate the method, or repeat a finding. An abstract goes from the problem and its challenge to the approach and findings, and a conclusion stays short.
• RULE-05 Dying metaphors: Delete clichés like "pushes the boundaries", "paradigm shift", or "state of the art".
• RULE-06 Plain English: Prefer "use" over "leverage", "method" over "methodology", "feature" over "functionality".
• RULE-07 Positive form: Prefer "trivial" to "not important", "forgot" to "did not remember"; do not stage "X, not Y" antithesis ("not just X, but Y") for emphasis.
• RULE-08 Claim calibration: Calibrate verbs to evidence; do not write "proves" when the evidence is "suggests"; keep qualifiers that change the claim, and cut caveat tails that add nothing.
• RULE-09 Parallel structure: Express coordinate ideas in the same grammatical form.
• RULE-10 Related words together: Keep subject close to verb and modifier close to modified; split long parentheticals.
• RULE-11 Stress position: Place new or important information at the end of the sentence.
• RULE-12 Long sentences: Vary sentence length; split a sentence when the reader must hold more than one open clause to reach its point, and let a long sentence stay if it carries one connected thought.

Field-observed rules (maintainer observation of LLM output, 2022-2026):

• RULE-A Bullet overuse: Keep prose in paragraphs when ideas connect; bullets only for genuine lists; avoid forced 3-item triads; table cells hold numbers or short labels.
• RULE-B Dash overuse: Do not use em or en dashes as casual sentence punctuation; prefer commas, colons, parentheses, or a new sentence; keep semicolons rare.
• RULE-C Same-starts: Do not open two or more consecutive sentences with the same word.
• RULE-D Transitions: Do not open sentences with additive-only transitions ("Additionally", "Furthermore", "In addition"), and let "Moreover" or "Notably" open at most one sentence in a paragraph; keep words that mark a contrast, cause, or sequence the text already establishes ("However", "Thus", "First"), and never add a cause just to join sentences.
• RULE-E Summary closers: Do not end every paragraph with a sentence that restates its point.
• RULE-F Term consistency: Once you define a term or abbreviation, keep using it; do not alternate synonyms.
• RULE-G Title case: Use title case for section and subsection headings; articles and short prepositions stay lowercase. A full-sentence bold run-in lead takes sentence case.
• RULE-H Citation discipline (critical): Support factual claims with verifiable citation or concrete evidence; never fabricate citations.
• RULE-I Contractions: Prefer "it is" / "does not" / "cannot" over "it's" / "doesn't" / "can't" in formal technical prose.

Escape Hatch

"Break any of these rules sooner than say anything outright barbarous." — George Orwell, "Politics and the English Language" (1946), Rule 6. Most rules are defaults to apply with judgment. Depart when the sentence reads better for it, not to vary the pattern. A few are firm (do not invent citations, evidence, or numbers; no placeholders in finished text). Judge a draft by its overall pattern: one that obeys every default in every sentence reads as machine-made.

House Style

The rules start from classic guides and are tuned to one author, Yue Zhao. A House Style section in RULES.md lists that author's habits, drawn from first-author ML papers written before AI writing tools: the moves of an abstract, introduction, results paragraph, and conclusion; about one sentence in four past 30 words, carrying parenthetical glosses; and sparing signposts. Where a style rule conflicts with a habit there, follow the habit, but never against the facts or the firm rules. Read that section in RULES.md before drafting a paper or proposal section.

Full Rule Bodies (Canonical)

Full directive text, BAD/GOOD example pairs, and rationale per rule: see .agent-style/RULES.md in this project, or https://raw.githubusercontent.com/yzhao062/agent-style/v0.9.0/RULES.md for the pinned canonical source.