Style code and local writing

WriteLikeMe SVGs include a versioned JSON profile in their <metadata> element. This lets the style code be shared as a file and read back into the package. The current schema is v3, which carries the seven visible categories plus voice cues, habit signals, grammar mistake signals, sentence-structure signals, readability metrics, style classification, domain detection, frequent words, and an optional named context string (email / blog / docs ...) stored in the SVG metadata. v1 and v2 style codes are automatically upgraded to v3 when read.

Read a profile from an SVG

const profile = WriteLikeMe.readStyleCode(svg);

readStyleCode(svg) returns an object with schemaVersion, the seven visible categories, additional signals, frequentWords, and available count stats. For SVGs created before profile metadata was added, it estimates the seven category scores from the visible marks; other signal values are not recoverable.

Get a style brief for an AI model

const brief = WriteLikeMe.describeStyle(profile); // or describeStyle(svgString)

describeStyle returns about seven short lines (register, sentences, voice, habits, punctuation, reading level, signature words), roughly 100 tokens. Paste it into a prompt so a language model writes in that voice; this works far better than the template writer below. The CLI equivalent is writelike-me brief --style-code me.svg, and the MCP tool is describe_style (see AI harnesses).

Example output:

Register: casual and conversational; warm, positive tone.
Sentences: short, with varied lengths; simple structure and everyday words.
Voice: speak in first person (I/we); use contractions.
Habits: ask the reader questions; use exclamation marks.
Punctuation: balanced commas and full stops.
Signature words: honestly, ship, team.

Write locally from a prompt

const draft = WriteLikeMe.writeLikeMe(
  'Write a short paragraph about community gardens.',
  profile
);

writeLikeMe(prompt, styleCode) accepts a profile object or a WriteLikeMe SVG string and returns a deterministic, template-based draft. The optional styleCode defaults to a neutral profile. The local writer uses the prompt topic and measured style cues; it makes no network requests and requires no API key. It does not have language-model reasoning, cannot reliably follow complex instructions, and should be treated as a basic writing aid rather than an exact imitation.

The CLI accepts the same SVG profile:

writelike-me write --style-code style.svg "Write a short paragraph about community gardens."

The SVG metadata contains profile scores and up to five frequent words, not the original writing sample. Frequent words can still reveal personal information, so inspect the profile before sharing it. AI providers do not automatically read SVG metadata; an application that uses an external model must call readStyleCode and pass the resulting JSON as explicit context.

checkDraft(profile, draftText, options)

Scores a draft against a style profile and returns { score, pass, issues }:

• score (0-1): heuristic similarity between the draft and the profile, computed with the same comparison math as compareTextStyle.
• pass: true when score >= options.min (default 0.8).
• issues: at most five { signal, yours, draft, fix } entries, largest gaps first, only gaps of 0.15 or more. Each fix is one short actionable sentence.

It covers sentence length, formality, complexity, contraction rate, first and second person, punctuation, question/exclamation rate, tone and more. The profile argument accepts a profile object or a WriteLikeMe SVG string. The CLI equivalent is writelike-me check; the MCP tool is check_draft.