---
title: Overview
description: lawlint flags the patterns that make legal and general prose sound machine-written, and suggests more human, direct alternatives.
icon: book-open
sidebar:
  order: 1
---

lawlint reads a document and returns a focused list of things worth revising —
the hedges that hedge nothing, the tricolons, the throat-clearing openers, the
em dashes doing a comma's job — plus a **human-likeness score** from 0 to 100.

It is a native binary. Linting needs no configuration, no network, and no
account: the deterministic rules run entirely offline over text you already
have on disk.

```sh
lawlint contract.docx
```

## Start here

<CardGroup cols={3}>
  <Card title="Quickstart" href="/docs/quickstart" icon="rocket">
    Install the CLI, lint a document, read the output.
  </Card>
  <Card title="Rule reference" href="/docs/rules" icon="list">
    Every built-in rule, what it catches, and why.
  </Card>
  <Card title="CLI reference" href="/docs/reference/cli" icon="terminal">
    Every command, flag, and exit code.
  </Card>
</CardGroup>

Nothing to install yet? The <a href="/playground">playground</a> runs the same
engine in the browser, compiled to WebAssembly — text typed there never leaves
the page.

## What it checks

Rules come in two kinds, and the distinction decides whether a check needs a
model.

<Columns cols={2}>
  <Column>
    **Hard rules** are deterministic. Phrase, leading, density, and statistical
    checks over the text itself — no model, no network, no variance between
    runs. Every rule that ships enabled by default is a hard rule.
  </Column>
  <Column>
    **Soft rules** are inferential: judgements a regular expression cannot make,
    evaluated by an optional AI judge from a written rubric. They run only when
    you ask for them with `--judge`.
  </Column>
</Columns>

See [Hard and soft rules](/docs/concepts/rules) for how the two kinds map onto
the engine's `static`, `statistical`, and `inferential` tiers.

## What the score means

The human-likeness score aggregates only **detection**-intent rules — the
subset corpus-validated to distinguish AI-generated from authentic human legal
prose. **Style**-intent rules (the Oxford comma, semicolons, sentence length,
the Orwell rules) still report findings and still participate in `--fix`, but
they never move the score.

Because the score is normalized per 1,000 words, it reflects issue *density*
rather than document length. [Scoring](/docs/concepts/scoring) has the formula.

## Beyond linting

<CardGroup cols={2}>
  <Card title="Fixes and tracked changes" href="/docs/guides/fixing" icon="wand">
    `--fix` rewrites what is safe to rewrite. In Word, as native tracked changes
    with a review comment per fix.
  </Card>
  <Card title="The AI judge" href="/docs/guides/judge" icon="scale">
    Add inferential findings that deterministic rules cannot reach.
  </Card>
  <Card title="Learn your voice" href="/docs/guides/learn" icon="graduation-cap">
    Mine a personal rule package from writing you have already done.
  </Card>
  <Card title="Author a rule" href="/docs/guides/authoring-rules" icon="file-pen">
    Rules are Markdown files with YAML frontmatter. Write your own.
  </Card>
</CardGroup>

## Privacy

Linting is local and offline. Nothing is uploaded, and the installers do not
phone home with your documents.

The optional AI features are the exception, and they are opt-in twice over:
they do nothing until `lawlint init` configures a model, and — with a hosted
provider — they send the text being judged to that provider. API keys live in a
user-level credential file, never in your project.