Typography

Typography

Announce at start: "I'm using the humane:typography skill to separate safe typographic corrections from editorial choices."

Polish text without taking authorship away from its author. This skill owns mechanical microtypography and an editorial/spelling review. It may apply only corrections whose meaning, voice, and cadence do not change; everything else is a proposal.

Boundary

Do not rewrite for concision, vary repeated words, remove sentence fragments, or reduce dash frequency under the label “cleanup.” Those are authorial choices.

Establish the contract

Before recommending a rule that differs by language or standard, establish:

1. Language: English, Russian, or named spans of each in mixed text.
2. Standard: the project's house style if one exists; otherwise the named editorial/typographic standard (for example Chicago or New Hart's Rules for English; Lopatin, Milchin, or a publisher's house style for Russian).
3. Medium: plain text, Markdown, interface string, HTML source, or print.

Read nearby content, localization metadata, and the project style guide first. If language or standard is unstated and the answer would materially change a recommendation, ask the user before making that recommendation. Do not silently choose US vs British spelling, quote style, serial-comma policy, Russian ё policy, dash convention, or how aggressively short words are protected. If the distinction cannot affect the requested mechanical check, continue and state the narrow assumption instead of asking ceremonially.

For mixed-language text, apply each language's rules only inside the span whose language is known. Never normalize a quoted title, command, URL, code span, identifier, captured user quote, or brand spelling merely because it looks foreign.

Quick reference

Read only the language reference that applies. Read both for mixed-language text, and label the boundaries used.

Pass 1 — Mechanical corrections

Inspect the smallest complete source that preserves context. Protect code, URLs, file paths, identifiers, Markdown destinations, placeholders, template syntax, and verbatim quotations from automated replacement.

Classify a change as mechanical only when there is one valid correction under the established contract and no plausible shift in meaning or voice. Typical mechanical work includes:

• required nonbreaking spaces and protected pairs;
• spaces around punctuation and symbols under the chosen language standard;
• unambiguous quote, apostrophe, dash, ellipsis, multiplication-sign, degree, number/unit, №, and § corrections;
• invisible or damaged Unicode such as zero-width residue, a replacement character, or a literal escaped Unicode sequence accidentally shown to a reader;
• orphan Markdown emphasis markers or duplicated punctuation only when the intended character is unambiguous.

The bundled helper provides a repeatable first pass and never writes source files:

python3 scripts/typography.py check --language en --standard chicago path/to/text.md
python3 scripts/typography.py format --language ru --standard lopatin path/to/text.md

Use - for stdin. check emits a structured report; format emits corrected text to stdout. Review its output before applying a patch. The helper is a conservative detector, not an editorial authority: a clean run does not prove the prose is correct. Its deliberately closed standard list is chicago, new-harts, or ap for English and lopatin, milchin, or gost for Russian; apply a project's custom house style in the human review rather than inventing a CLI token for it.

Pass 2 — Editorial and spelling review

Read editorial-pass.md. Report these separately from the mechanical pass:

• spelling that depends on locale, house dictionary, ё, a proper name, or product terminology;
• punctuation with more than one defensible reading;
• sentence rhythm, fragments, repetition, paragraphing, tone, and dash density;
• suspicious LLM habits that may be intentional: symmetrical three-part lists, generic transitions, excessive headings, rhetorical contrasts, emoji, arrows, or repeated em dashes.

Offer a small set of concrete alternatives and explain the effect of each. Never apply one until the user chooses it or has already authorized that class of editorial change.

Output contract

Always begin with Context: language, standard, medium, protected regions, and assumptions.

For a formatting or apply request, keep Mechanical corrections (Location | Before | After | Rule) and Editorial choices (Location | Current | Options | Decision needed) in separate sections. “None” is valid; do not invent choices to fill either table. End with Verification: helper command run, sources inspected, and anything not verified.

For a standalone review, use the repository's shared review contract:

1. Findings — one table with Severity | Location | Before | After | Why and an added Category column whose value is Mechanical or Editorial choice. The category keeps the two classes separate without producing two reports. For an editorial choice, After contains the options, not a silent rewrite. Cap at 15 and never pad.
2. Considered but Rejected — two to five real candidates and why they were left alone.
3. Verification — what was actually checked; name any missing dictionary, house guide, language boundary, or runtime rendering check.
4. Verdict — exactly Block, Needs changes, or Approve.

Under humane:review, return only findings in the orchestrator's required shape; its severity scale, consolidation, cap, and verdict win.

When the user asked for a review, remain read-only. When the user asked to apply fixes, apply the mechanical section only, then pause on unresolved authorial choices. Preserve the original file format and line endings.

Safety and honesty

• A nonbreaking-space policy is not universal. Protect only pairs justified by the selected language, standard, and medium.
• “Looks AI-generated” is not a defect and cannot be inferred from punctuation. Report only the observable pattern and its reading cost.
• Never claim a full spelling check without the relevant dictionary or house word list. Name what was checked manually.
• Never normalize code-like content or external quotations without explicit authorization.
• The user's explicit words > the selected standard > the project's existing system > this skill's defaults > personal taste. Surface conflicts.