writinglint-core

The engine behind WritingLint: a Document model over a dependency-parse + POS graph, an authorable Rule API (defineRule), config resolution, and the Linter. Bring your own parser (e.g. writinglint-parser-node).

import { Linter, resolveConfig } from 'writinglint-core';
import { loadParser } from 'writinglint-parser-node';
import { recommended } from 'writinglint-rulepack-ai-style';

const linter = new Linter(await loadParser({ modelDir: './models/xsmall' }));
const { lints } = await linter.lint('Trust the graph, not the vibes.', resolveConfig(recommended));

See the docs for the full API.

Building language and standards rulepacks

WritingLint keeps reusable analysis in core and policy decisions in rulepacks. The parser can identify its BCP 47 languages, model version, fingerprint, and available annotations through ParserDescriptor. A rule can declare the capabilities it requires; the linter records the rule as skipped when the active parser cannot supply them.

Callers can give Linter.lint() structured source regions, independent span annotations, and services without changing the parser:

const report = await linter.lint(text, config, {
  language: 'en',
  regions: [
    { id: 'step-1', role: 'step', mode: 'procedural', start: 0, end: 24 },
  ],
  annotations: [
    { kind: 'measurement', provider: 'units-v1', start: 8, end: 13 },
  ],
  services: { terminology },
});

Core provides these standard-neutral extension points:

• DocumentRegion for headings, lists, procedures, steps, notes, warnings, cautions, tables, quotations, code, and custom roles;
• SpanAnnotation for proper names, measurements, abbreviations, terms, and other independently produced annotations;
• ParserDescriptor and RuleMeta.requires for explicit capability checks;
• optional token lemmas, Universal Dependencies morphological features, and calibrated parser confidence;
• TerminologyProvider, InMemoryTerminologyProvider, and LayeredTerminologyProvider for standard, industry, organization, project, and document vocabularies;
• CountPolicy and countSentenceUnits() for inspectable, standard-specific word counting; and
• structured finding evidence, split-point and other positional anchors, multi-dimensional magnitude metrics, visible assumptions, and per-rule execution records.

The core does not decide whether a word, sentence, or construction conforms to a standard. A rulepack combines these capabilities and remains responsible for that interpretation.