no-em-dashes

A Claude Code skill that produces writing
with zero em dashes, and, just as important, writing built out of sentences that never
reach for one in the first place.

Heavy em-dash use has become one of the most recognizable fingerprints of
machine-written text. Many readers now skim a dash-laden paragraph and think "a bot
wrote this." This skill keeps prose reading as a person's, and it forces tighter,
more deliberate sentences along the way.

Built and shared in the open by Exmergo, the team building
AI agents for data analytics. See more of our work at
exmergo.com/open-research.

Install (Claude Code)

Option 1: As a plugin (recommended)

This skill is published through the Exmergo
marketplace. Add the catalog once, then install the skill. Inside Claude Code:

1. Add the marketplace

/plugin marketplace add exmergo/exmergo-agent-plugins

1. Install the skill as a plugin

/plugin install no-em-dashes@exmergo

The skill then loads automatically and runs whenever Claude writes or edits prose for
you. To update later, run /plugin marketplace update exmergo.

Option 2: Manual copy

If you would rather not use the plugin system, copy the skill folder into your skills
directory:

git clone https://github.com/exmergo/skill-no-em-dashes
cp -r skill-no-em-dashes/skills/no-em-dashes ~/.claude/skills/   # personal, every project
# or, project-scoped:
cp -r skill-no-em-dashes/skills/no-em-dashes .claude/skills/

Claude Code auto-discovers any folder containing a SKILL.md. Either way, you do not
have to mention em dashes; clean output is the default.

Use with other coding agents

The two-command install above is specific to Claude Code, which has a plugin
marketplace. Most other agents do not install packages; they read committed
instruction files. The good news is that one file reaches almost all of them.

Start here: AGENTS.md

This repo ships an AGENTS.md, the open standard (now under the Linux
Foundation) that Cursor, GitHub Copilot agent mode, Codex, Continue, Aider, Windsurf,
and others read automatically. To get the behavior in your own project, copy the
contents of our AGENTS.md into your repository's AGENTS.md (merge if you already
have one). That single step covers the majority of agents, with no install command.

AGENTS.md is the canonical source of the guidance. The per-tool files below are thin
pointers to it, so nothing drifts.

Per-tool files

Two honest caveats

• Mostly project-scoped. For these agents you get the rule by placing the file in a
  given project, not by installing it once for everything. The cleanest "applies
  everywhere" options remain the Claude Code plugin above and Cursor's Remote Rules.
• The checker needs a shell. The prose guidance works in every agent. The
  check_em_dashes.py scan only helps in agents that can run a terminal command (Cursor
  agent mode, Windsurf, Claude Code), and only when the rule tells them to run it.

What it does

When the skill is active, any prose Claude writes or edits comes out free of:

• the em dash itself (U+2014),
• and its disguises: a double hyphen (word--word or --), a hyphen padded with
  spaces (-), a stray en dash between words, and the horizontal bar.

The point is not to swap one dash character for a look-alike. It is to recognize the
sentence shape that wanted the dash, then rewrite with real punctuation: a comma,
semicolon, colon, period, or parentheses. The skill includes a playbook for each job
an em dash typically does (the dramatic reveal, the "not just X but Y" pivot, the
interruptive aside, joined clauses, the summary dash, and interrupted dialogue) with
worked before-and-after examples.

It deliberately leaves legitimate hyphens alone: compounds like well-known and
state-of-the-art, number and date ranges like 10-20 and 2019-2024, and hyphens
in code, file paths, URLs, and math.

Repository layout

AGENTS.md                          canonical cross-agent guidance (see "Use with other coding agents")
.claude-plugin/
  plugin.json                      Claude Code plugin manifest (name, version, author)
.cursor/rules/no-em-dashes.mdc     Cursor rule, points to AGENTS.md
.github/copilot-instructions.md    GitHub Copilot instructions, point to AGENTS.md
.windsurf/rules/no-em-dashes.md    Windsurf rule, points to AGENTS.md
skills/
  no-em-dashes/                    the Claude Code skill itself
    SKILL.md                       the instructions Claude Code loads
    scripts/check_em_dashes.py     a checker that flags em dashes and stand-ins
    evals/                         test prompts (evals.json) and recorded results
      results/em_dash_counts.csv   em dash counts per eval, with skill vs without
tests/                             standard-library unit tests for the checker
MAINTAINING.md                     release flow, private-to-public, and sync notes
CONTRIBUTING.md  CODE_OF_CONDUCT.md  LICENSE

The writing guidance lives in two self-contained files on purpose: SKILL.md for
Claude Code and AGENTS.md for every other agent. They mirror the same rule because
each ships independently; see MAINTAINING.md for how they stay in sync.

The checker

The bundled checker is a single dependency-free Python script. Use it on a file or
through standard input:

python skills/no-em-dashes/scripts/check_em_dashes.py path/to/draft.txt
echo "Your text here" | python skills/no-em-dashes/scripts/check_em_dashes.py

It prints every hit with its line, column, and a suggested fix, and it exits with a
nonzero status if anything is found, so you can wire it into a pre-commit hook or CI
step. It skips fenced code blocks, inline code, and URLs so it does not flag the
hyphens that belong there.

Does it work?

Yes. We ran five realistic prompts twice each, once with the skill active and once with
a skill-free baseline, then counted the em dashes in every output.

A note on method: the baseline runs are generated with the skill fully disabled. A
skill that is installed globally will otherwise clean the baseline too and hide the
effect, so isolating it matters.

The skill held every output at zero. The skill-free baseline produced em dashes in four
of the five prompts, and the count climbed with length and rhetorical heat, peaking at
18 in the landing page. Short, punchy copy was the only baseline that stayed clean,
which fits the pattern: em dashes surface most in longer, reflective, or persuasive
prose.

The prompts live in evals.json, the per-eval
counts in em_dash_counts.csv,
and example outputs under
evals/results/examples/. You can also
explore the results in an interactive chart on Exmergo Viz,
our AI agent for charting and dashboards.

Running the tests

The checker has a standard-library test suite (no dependencies):

python -m unittest discover tests

Contributing

Contributions are welcome. See CONTRIBUTING.md for how to propose
changes, and please keep your own prose and commit messages em-dash free (the checker
will tell you if you slip).

Privacy

The skill collects no data and runs locally. See PRIVACY.md.

License

MIT.