demdash

Remove em-dashes from LLM-generated content in a grammatically correct manner

LLMs love em-dashes—they use them constantly—sometimes multiple times in a single sentence. This can make AI-generated content feel repetitive—and stylistically monotonous.

demdash solves this—quite elegantly, if I say so myself—by using an LLM to intelligently rewrite text—replacing em-dashes with more varied punctuation—while preserving the original meaning.

-What’s striking isn’t the complexity—it’s the pauses—the little hesitations where certainty should live—where context leaks out and assumptions rush in to fill the gap. You can feel it when you read the docs—when a clause trails off—when an em dash does the work of a missing decision—signaling thoughtfulness—while quietly avoiding commitment.
+What’s striking isn’t the complexity; it’s the pauses, the little hesitations where certainty should live, where context leaks out and assumptions rush in to fill the gap. You can feel it when you read the docs, when a clause trails off, when a parenthetical or colon does the work of a missing decision, signaling thoughtfulness while quietly avoiding commitment.

Installation

npm install demdash
# or
bun add demdash

Quick Start

import { demdash } from 'demdash';

const text = "The package is simple—just one function—and it works great.";
const result = await demdash(text);

console.log(result);
// "The package is simple (just one function) and it works great."

Try It Now

export OPENAI_API_KEY=your-key-here
bunx demdash --example

This picks a random sample text and shows before/after.

CLI Usage

Process a file:

bunx demdash input.txt > output.txt

Process piped input:

echo "Text with em-dashes—like this—gets fixed." | bunx demdash

Use a specific model:

bunx demdash --model gpt-4o input.txt

View help:

bunx demdash --help

Configuration

Environment Variables

Variable	Description	Default
OPENAI_API_KEY	OpenAI API key (required if not passed in options)	-
OPENAI_BASE_URL	Custom API base URL	OpenAI default
DEMDASH_MODEL	Model to use. gpt-5 is more nuanced	gpt-4o-mini
DEMDASH_MAX_TURNS	Maximum LLM calls per invocation	3

Options

interface DemdashOptions {
  apiKey?: string;        // OpenAI API key
  model?: string;         // Model to use (default: 'gpt-4o-mini')
  maxTurns?: number;      // Max retry attempts (default: 3, max: 10)
  baseURL?: string;       // Custom API base URL
  provider?: LLMProvider; // Custom LLM provider
}

Options take precedence over environment variables.

Example with Options

const result = await demdash(text, {
  apiKey: process.env.MY_OPENAI_KEY,
  model: 'gpt-4o',
  maxTurns: 5,
});

Custom Providers

Use any LLM by implementing the LLMProvider interface:

import { demdash, LLMProvider } from 'demdash';

const myProvider: LLMProvider = {
  async complete(prompt: string): Promise<string> {
    // Call your preferred LLM
    const response = await myLLMClient.generate(prompt);
    return response.text;
  }
};

const result = await demdash(text, { provider: myProvider });

How It Works

1. Detection: Checks if the input contains em-dashes (—), en-dashes (–), or double-hyphens (--)
2. Prompt Construction: Builds a prompt instructing the LLM to replace dashes with alternative punctuation (commas, parentheses, colons, or periods)
3. LLM Call: Sends the prompt to the configured LLM
4. Verification: Checks if dashes remain; if so, retries up to maxTurns times
5. Result: Returns the cleaned text

Why?

Look, I personally love em-dashes. I've used them heavily since my time building Medium—they're elegant, expressive, and versatile. But em-dashes have become a telltale sign of AI-generated content, and not everyone shares my enthusiasm for them.

In case it’s not obvious, this library is a joke. But if you genuinely need to de-AI your text, here you go.

As McSweeney's put it:

So next time you read something and think, "AI wrote this—it has a lot of em dashes," ask yourself: Is it AI? Or is it just a poet trying to give you vertigo in four lines or fewer?

License

MIT