Strunk & White for the prompt-engineering era

What it does

agent-style is a drop-in ruleset for AI coding and writing agents—Claude Code, Codex, Cursor, and others—that loads at generation time to shape how they draft technical prose. It covers API docs, design docs, research papers, READMEs, and commit messages, but deliberately skips fiction, marketing, and anything where rhythm matters more than precision. The rules are not a post-hoc linter; they sit in the agent’s context window and attempt to prevent violations before the first token is written.

The interesting bit

The rules split evenly between canonical authorities—Strunk & White, Orwell, Pinker, Gopen & Swan, each cited by chapter and section—and nine field-observed patterns the author logged from AI output between 2022 and 2026. That hybrid approach is the project’s core bet: classic style guides catch the curse of knowledge, but only empirical observation catches LLM-specific tics like over-bulleting, em-dash abuse, and summary-sentence bloat.

Key highlights

• 12 canonical rules with verified citations, plus 9 field-observed rules targeting LLM failure modes like citation hand-waving and transition-word spam.
• A sanity benchmark on v0.3.0 shows directional drops in “AI-tell” violations: Claude Opus 4.7 and GPT-5.4 each fell roughly 45%, while Gemini 3 Flash fell 82% across 10 fixed prose tasks.
• Two enforcement modes: soft enforcement at generation time, and an opt-in style-review skill that audits a finished draft without touching the source file.
• Explicit scope boundaries: in scope for technical docs and papers, out of scope for fiction, poetry, marketing copy, and non-English prose.
• An Orwell escape hatch baked in: “Break any of these rules sooner than say anything outright barbarous.”

Caveats

• The benchmark numbers are explicitly directional, not statistically significant: only 10 tasks × 2 generations per condition.
• GitHub Copilot CLI’s programmatic mode does not load instruction files, so it saw +3% noise and was excluded from the published figure.
• The field-observed rules are transparently attributed as the author’s own logged patterns, not literature-backed.

Verdict

Worth a look if you maintain agent configurations for technical teams and are tired of editing out LLM-isms by hand. Skip it if you need creative writing, marketing flair, or statistically rigorous proof that style guides change model behavior.