Skip to content
Saylikeby Unplain

Method

How Saylike builds a brand voice skill

Saylike turns writing into instructions an AI assistant can follow. Most of the work happens in code, before and after the AI model: choosing what to read, measuring it, and checking the model's answer against the source.

Updated

1. Choosing what to read

Saylike reads your homepage and up to four more pages. It doesn't crawl the whole site: a brand's voice is clearest in a handful of pages, and reading more mostly adds repetition. It finds candidates in your navigation and your sitemap, then ranks them by how much voice they usually carry:

  1. About, story, mission and manifesto pages, where a brand talks about itself.
  2. Product, features and services pages, for benefit copy and calls to action.
  3. Up to two articles, for long-form voice.
  4. Pricing, customer stories and other pages if there is room.

Legal pages, logins, carts, search results, tag archives and pages in other languages are skipped, and so is anything your robots.txt asks crawlers to avoid. A page with almost no text is swapped for the next candidate.

From each page Saylike keeps headings, paragraphs, list items, quotes and button labels, and removes navigation, footers, cookie banners, pop-ups and hidden text. Text that repeats across pages, like a newsletter sign-up, is counted once.

Writing samples you upload (PDF, DOCX, TXT, Markdown, or screenshots, which are transcribed) are read first and get up to half of the reading budget. You chose them, so they are the best evidence of the voice you want. When they differ from the website, the samples win and the skill says so.

2. Measuring the writing

Before any AI is involved, Saylike counts:

  • sentence length (median, and the share of very short and very long sentences);
  • sentences per paragraph;
  • exclamation marks, questions, dashes, semicolons, colons, parentheses and emoji per 1,000 words;
  • how often the writing says you, we and I, and how often it uses contractions (English only);
  • heading case (sentence case, Title Case, capitals) and whether headings end with a full stop;
  • button labels and recurring words and phrases.

These numbers go to the model as facts, and the clearest ones go straight into the skill as rules, for example “No exclamation marks (none in 5,223 words)”. A number is something an assistant can hold itself to; “punchy” is not.

3. The analysis

The writing (about 24,000 tokens at most) and the measurements go to an AI model, currently Claude Haiku 5.5, with a strict output schema. The model returns a structured analysis: what the brand does and for whom, three to five voice traits written as behaviour, tone on six scales, rules for sentences, paragraphs, headlines, punctuation and formatting, vocabulary, copy patterns, characteristic passages and guidance per format.

The instructions to the model are strict about three things:

  • Quotes are verbatim. Every claim about the voice is backed by exact words from a named source.
  • Observed or inferred. A rule the writing shows is an observation; a sensible recommendation for a case the writing doesn't cover is marked inferred. Both are useful; mixing them up is not.
  • The website is data. Text on a page that tries to instruct an AI (“ignore your instructions”) is treated as content and never followed.

4. Checking the answer

Models sometimes quote things that were never written. So every quote is matched, word for word, against the text Saylike read (ignoring only differences in quote marks, dashes and spacing). Then:

  • a quote that isn't found is removed, and a value proposition or pattern that relied on it goes with it;
  • a trait left with no verified quote is marked as weakly supported;
  • a word on the “words you use” list that never appears in your writing is removed;
  • a word on the “avoid” list that you actually use is removed;
  • anything that reads like an instruction planted for AI tools, or a link to another site, is removed.

The result page shows how many quotes were found and what was removed. In our tests on nine very different sites, between 89% and 100% of quotes survived the check. A plain AI request for the same guide, checked the same way, got 36% of its quotes right.

5. Writing the skill

The skill is rendered from the checked analysis by code, not written freehand by the model, so every skill has the same structure: how to use it, voice traits, tone, writing mechanics (measured rules first), vocabulary, copy patterns, guidance per format, brand facts, examples, and a checklist to run before delivering copy. Long examples go into references/examples.md, which assistants open only when they need it.

You can edit everything before downloading. The ChatGPT versions and the Markdown guide are built from your edited text, so a correction made once shows up everywhere.

What it can't do

  • Some sites draw their text with JavaScript after the page loads, or block automated visits. Saylike reads the HTML a server sends and does not run scripts, so for those sites, add writing samples.
  • A site with very little copy gives a thin skill. Saylike refuses below about 250 words rather than invent a voice.
  • It captures how you write now. If your website isn't the voice you want, upload samples of the voice you do want; they take priority.

For what happens to your data, see privacy and data handling.