Help / Building on sitelaunch

Reading this help center with a model

Every page as Markdown, an llms.txt index, one search endpoint, and the same search over MCP.

The pages here are written for a person, and served in four shapes so a model does not have to scrape one.

One page, as Markdown

Put .md on the end of any page address.

/help/merchant/take-a-booking       the page
/help/merchant/take-a-booking.md    the same page, as Markdown

Drawings are HTML in the page's own source. In the Markdown mirror each one is replaced by its caption, on a line that begins > On screen:, so nothing about a picture reaches you as markup you have to parse.

Every page, indexed

/help/llms.txt        every page named, linked and described
/help/llms-full.txt   every page inlined, for a model with room for all of it

llms.txt follows the convention: a title, a line on what sitelaunch is, then one bullet per page pointing at the .md. It also names the machine surfaces at the bottom, including the MCP.

Search

GET /help/search.json?q=deposit&limit=12

It answers with the query and an array of hits. A hit is a section, not a page:

• slug and persona, which together address the page,
• title, the page's own title,
• heading, the section the match was in,
• id, that section's anchor,
• url, the whole address including the anchor,
• snippet, the matching words in context,
• score.

Sections rather than pages is deliberate. "Which page" leaves a reader to find the paragraph, and on a page of eight numbered steps that is most of the work still to do.

A word has to be in the section's text to count at all, and that is worth one. Being in the page title adds eight and in the heading four, so those stack. A whole-word match adds two, because "pay" should not rank "payments" above "pay". The query as a phrase adds six.

Two rules matter more than the weights. Every word has to match, not any, so "take a booking" does not return every page containing "take". And at most two sections come back from one page, so a thorough page cannot fill the results and hide the page that actually answers.

limit defaults to 12 and is capped at 25.

The same search, over MCP

The MCP server carries three help tools, all on the store:read scope:

• search_help for a question,
• read_help for one page,
• list_help for what exists.

They call the endpoint above rather than keeping their own copy, so a merchant typing "deposit" into their console and an assistant asking the same thing get the same sections back. One implementation, several clients.

The store itself

This center is documentation. To read or change an actual store, that is the MCP:

/mcp           the server
/mcp/guide     the tools, the scopes, and the rules

The rule worth knowing before you write anything: words, looks and landers are a draft until publish_store, and products are not. A product is live on an open store the moment it is added. Every write says which it was in its answer.

What is stable

The addresses above and the shape of a search hit. The pages themselves are edited as the product changes, and each one carries its own reading time and what it assumes you have already done.

Still stuck? Ask us. Models: this page is Markdown, and every page is in llms.txt.