Dashes

The dashes rule sets the parenthetical dash in the form each locale prescribes, and leaves compound hyphens, numbers, ranges and arrows alone.

Updated 21 October 2026 · spec 1.12.0

The dashes rule renders the parenthetical dash, the one that interrupts a sentence, in the form the locale prescribes. A spaced hyphen or a doubled hyphen becomes the locale’s dash: a tight em dash in en-US, a spaced en dash in de-DE and en-GB, a spaced em dash in fr and ru. An em dash or en dash the author already typed is corrected in the same way, in spacing and in length.

It is on by default, runs after spaces, ellipsis and ranges, and works in all four modes. It is deliberately conservative. It only touches a stroke that is symmetrically spaced and has text on both sides, and it never touches a hyphen inside a word, a number, a date, a command-line flag or a list marker.

How it works #

The rule scans the text for a dash token: a maximal run of dash characters. The characters that count are the hyphen-minus (U+002D), the hyphen (U+2010), the en dash (U+2013), the em dash (U+2014) and the minus sign (U+2212). A run is one token, so -- is one dash, not two.

Each token then goes through a series of checks. The first one that fails leaves the token exactly as written.

1. Length. A run of four or more is decoration (a horizontal rule, a signature line, a Markdown ----) and is skipped. Runs of one, two or three are candidates.
2. Symmetry. The token must have a single ordinary space on both sides, or on neither. A stroke that is spaced on one side only is not a dash: --force, -v, - item, -- signature, ->.
3. Content on both sides, on the same line. A dash at the start or end of a line has nothing to interrupt, which protects list items and thematic breaks.
4. Isolation. If the character beyond the spacing is another dash, a space or a no-break space, the token is skipped. This covers a- - a and any dash that touches a no-break space.
5. Clusters. A stretch of digits and dashes with two or more dash runs in it (2026-08-15, 212-555-1234, a—0–0) is left entirely alone.
6. Compound words. A single hyphen with no space around it is a compound-word hyphen: well-known, e-mail, Jean-Luc.
7. Numbers. A dash with a digit on each side is not handled here at all; it belongs to the ranges rule.
8. Roman numerals. A tight dash between two Roman numerals is left alone.
9. A single authored en dash is never rewritten.

A token that passes everything is replaced according to the locale’s dash.parenthetical value, with any ordinary spaces around the old dash replaced too. The rule always writes an ordinary space (U+0020). Whether that space becomes a no-break space is decided later by nbsp.

Locale data #

The rule reads one field, dash.parenthetical, which has five possible values: a length (em or en) and a spacing (tight or spaced), or none.

none means the locale has no verified convention for the parenthetical dash. The rule emits nothing for it, not even a fallback. The range dash is a separate field read by ranges; the two can differ within one locale. Russian, for example, sets a spaced em dash in a sentence and a tight em dash in a range.

Parenthetical dashes #

The same input, run through three locales, shows the three active conventions. The English input uses a spaced hyphen, the German one a doubled hyphen with no spaces.

A tight locale removes the spaces around the dash; a spaced locale adds them when the input has none. In de-DE the doubled hyphen becomes a spaced en dash:

Dashes that are already typed #

An em dash or en dash in the input is not preserved as the author’s choice. It is classified and replaced like a hyphen would be, so both its length and its spacing are brought to the locale’s form. en-US turns a spaced em dash into a tight one; de-DE turns a tight em dash into a spaced en dash.

A mixed run of dash characters is one token and is treated the same way. Here -– is a run of two:

The U+2010 hyphen and the U+2212 minus sign are ordinary candidates too, since they are typewriter or OCR substitutes for a hyphen-minus.

The lone en dash is never rewritten #

One case is exempt. A single, pure U+2013 is left alone in every locale. An en dash and an em dash do different jobs (connection and range against interruption), and a lone en dash is taken to be a deliberate mark of the first kind. In en-US, where the parenthetical form is an em dash, a spaced en dash stays an en dash:

The exemption applies to a run of exactly one en dash. A run mixing it with another dash character is still converted, as in the a-–b example above.

Compound words, flags and list markers #

A hyphen with no space around it is part of a word, and the rule does not read it as a dash. This is why the symmetry and compound checks come first. The input below contains a compound, an e-mail address, an identifier with a digit suffix and a name, and none of them changes:

Asymmetric spacing keeps command-line flags, list items and signatures intact. A leading - at the start of a line, as in a Markdown or YAML list, has no content to its left, so it is untouched. A signature -- Iurii at the start of a line is untouched for the same reason. In the output every stroke stays as typed; the only edit is a no-break space before °C, added by nbsp.

Lines made only of hyphens, such as --- under a heading or as a thematic break, are skipped by the length and line-content checks:

Numbers, dates and ranges #

A dash with a digit on both sides is a range candidate and belongs to the ranges rule, never to dashes. This holds whether or not ranges is turned on. With default options, 5-10, 34-36, Figure 5-10 and an ISO date such as 2026-08-15 come out byte for byte as they went in:

Turn the ranges rule on and the page range becomes an en dash (bound with an invisible word joiner so it cannot wrap), while the date is still declined. See Ranges for the details.

Since spec 1.3.0 a range candidate can also have one closed-up symbol such as $ or % next to a number. $15 - $20 and 35%-50% are therefore ranges territory too, and dashes does not convert them:

The same applies to a dash between two numbers that is not a range, such as a phone number (212-555-1234) or an ISBN. They contain more than one dash run in a cluster of digits, so nothing in the cluster is touched.

A dash with a digit on one side only is a different case. Price--$50--drop has letters next to the dashes and is a normal parenthetical in en-US:

In a locale with a spaced form this is one place where the rule declines. A tight dash that would gain spaces is left alone when it sits against a digit run which has another dash on its far side, because inserting a space there could change how the neighbouring token reads on a second pass. In de-DE, Anstieg--50%--war is therefore left as typed, exactly as Anstieg--50--war is. A locale with a tight form never hits this guard.

Roman numerals #

A tight dash between two Roman numerals is a range already in its correct form, and in Russian it is a tight em dash. Converting it with the spaced parenthetical form would damage correct text, so the rule skips it:

In the output the dash is unchanged; the only edit is the no-break space that nbsp adds after the one-letter word. The guard declines and never converts. XV-XVII typed with a hyphen is also left alone, so a hyphenated Roman range keeps its hyphen. Only capital Roman letters count, so lower-case words such as mix—did are still normalised.

Spacing and no-break spaces #

Only the ordinary space (U+0020) counts as spacing around a dash. A no-break space (U+00A0) or narrow no-break space (U+202F) beside a dash makes the whole token inert, on either side. A no-break space is not something this rule put there, so it is not this rule’s to reinterpret. The practical consequence is that the rule is stable under the pipeline’s own output: nbsp may later turn the space before a Russian em dash into a no-break space, and a second run then leaves it alone.

A dash whose spacing is part no-break, such as one the author typed as mot + U+00A0 + - autre, is therefore never normalised.

Punctuation and brackets #

In a spaced locale the rule does not create a space that another rule would remove. A dash directly before ., ,, ;, :, !, ?, … or a closing bracket, or directly after an opening bracket, is left alone. A tight locale emits no spaces, so there is nothing to undo and the dash converts:

Dash-like characters that are never touched #

These characters are never candidates and never produced: the soft hyphen (U+00AD), the non-breaking hyphen (U+2011, which the hyphen rule writes), the figure dash (U+2012), the horizontal bar (U+2015) and the small and fullwidth forms U+FE58, U+FE63 and U+FF0D. Each has a purpose of its own, and the rule does not read a specialised dash as a typo. A candidate whose neighbour is one of them is skipped as well.

Modes and skipped regions #

In html, markdown and yaml modes the rule only sees prose. Code spans, fenced code, URLs and HTML attributes are removed by the mode before the rule runs. The rule itself has no notion of a URL.

In text mode the whole input is prose, so a doubled hyphen inside a URL in plain text is converted like any other. Choose a mode that matches the content if the text mixes prose and code.

Turning it off #

Set the rule off with the rules option. Nothing else changes, because disabling a rule removes it from the sequence without reordering the rest.

import { transform } from 'polytypo';

transform('The plan - if there is one - fails.', {
  locale: 'en-US',
  rules: { dashes: false },
});

What it does not do #

• It does not convert a range such as 1914-1918. That is ranges, which is off by default.
• It does not change a hyphen inside a word. The few forms where a hyphen must not break across lines belong to hyphen.
• It does not turn a leading - or — into a dialogue dash. A leading hyphen is a list marker in Markdown, and rewriting it would break list structure.
• It does not turn -> or => into arrows. They are source code more often than prose.
• It does not reformat dates or other digit chains.
• It does not recognise a named date range typed with a hyphen. 1 May - 3 June has letters beside the hyphen, and in en-US it becomes 1 May—3 June. Only an en dash the author already typed (1 May – 3 June) is protected. Recognising the hyphen form would need month names, which the locale data does not carry.
• It emits nothing for locales whose dash.parenthetical is none (Greek, Spanish, Turkish):

Related #

• Ranges, which handles digit-flanked dashes
• Hyphen, which binds hyphens in listed Russian and Ukrainian forms
• Spaces and Nbsp, the rules that run before and after
• Rule order
• Options

Normative text: spec/rules/dashes.md.