This lesson replaces the open-ended request to “optimise this page” with a brief built from audit data, and shows how to use Claude to rewrite content from it without losing facts, voice or rankings. It ends with the only proof that counts: running extraction again.
What this lesson covers
- The Entity Brief Generator
- Inputs from the audit
- A prompt structure the model follows
- Scaling one brief across pages
- Guardrails and what stays human
- Measuring a rewrite
The Entity Brief Generator turns one row of the audit into a brief a writer or a model can act on: the intended entity and its canonical name, the attributes the page is missing, the competing entities to reduce and the current text. The lesson walks through it in six steps and then explains the prompt structure behind it — context, task, constraints, output and review — and why the order matters. Anthropic’s own guidance that putting long documents before the question can improve response quality by up to 30 per cent is one reason the page text goes first and the instructions last.
The guardrails are explicit. A rewrite must not cost you anything: every number and date kept exactly, the site’s voice kept, headings that already rank left alone, and legal or medical wording untouched, because rewriting a reviewed sentence resets its review. Some pages never publish without a person — pages with factual claims, pages that rank well now and anything in health, finance, legal or safety — and everything else is spot-checked at one in five as a minimum. The lesson then shows how the same brief structure runs across many pages and closes with a checklist for success — the intended entity dominant, the mismatch cleared, nothing lost — checked by re-extracting against the frozen snapshot.
The argument running through the lesson is that entity rewriting is the highest-leverage automation in the course and the easiest to do damage with, and that these are the same fact. A model will smooth “46%” into “around half” without anyone noticing, so the review step is what makes the automation safe to run at all, not a formality to cut when volume rises.
This lesson continues Module 5, Remediation: Fixing Entity Representation, in The AI-powered Entity Audit. It suits teams using AI for content who need rewrites they can defend line by line.