How we write answers

Retrieval systems chunk on headings and score whether a chunk answers a question by itself. So we do not publish a wall of prose. The only way to add substance is to add a block: a stable kebab-case anchor, a real question, a 40–60 word answer that makes sense quoted alone, then optional detail and sources.

Last updated 13 August 2026. Policy context sits on /editorial-policy.

What is the chunk contract?

Each block has an id that never changes after publishing, because citations point at it. The question is phrased as something a person actually types. The lead answer cannot say "this page", "above" or "as mentioned", because those words break when the chunk travels without the rest of the article.

Detail is unbounded markdown: tables, lists, the argument. Sources are required for any disputable claim. FAQs are a separate, shorter list used only for FAQPage markup. Mixing the two is how sites get their structured data ignored.

How do dates stay honest?

lastReviewed on the page and lastmod in the sitemap are derived from the same frontmatter field. Static chrome pages use a shared date map so the sitemap cannot claim today when the visible byline says July. We move a date only when the content changed.

Markdown mirrors are generated from the same validated objects that render HTML, advertised with rel="alternate" type="text/markdown", and noindexed so they do not compete with the human page. The rewrite that serves /answers/x.md and the generator that builds that file have to ship together. An advertised mirror that 404s is worse than none.

Who should skip this method page?

Skip it if you wanted to know whether a CPD course is recognised, whether you can paste client data into ChatGPT, or which credential you actually need. Those questions have their own pages. This one is for people checking whether they should trust the others.

The working tools that apply the method for a reader are the answers library and the credential checker.