← back

How the engine works

The logic behind the content migration tool, top to bottom. Two things loop backward, that's how it gets smarter.

automated step AI step human sign-off gate learning loop (feeds backward)

Onboard onboarding.html

Capture the client once: old site URL, the new Webflow CMS fields, brand voice, and a couple of decisions.

in: a new clientout: client-config.json

▼

Scrape scrape-wordpress.mjs

Pull every page + product off the old WordPress/Woo site via its sitemaps. Never invents anything (blank stays blank).

in: old site URLout: clean content inventory (title, SEO, body, images, specs)

▼

Target schema

The new CMS shape: collections + fields, each tagged migrate / generate / manual. Read from the live Webflow build (best) or the scope doc.

Match & map build-match-map.mjs

Old → new URL map, 301 redirects, and orphans (old pages with no new home).

▼

Match sheet human gate build-match-sheet.mjs

Proposes which old page maps to which new collection (using the global learning below), then a dev or client confirms, reassigns, or overrides. Nothing proceeds until it's signed off.

out: match.confirmed.json

↑

Structure learning — global, all clients

Every reassignment/override is stored as a reusable pattern (not a URL) in one shared store. It improves the matching in step 3 for this build and every future client. Client #10 starts ahead of client #1.

↺ back to step 3

▼

Split & fill AI split-fill.mjs

Extract, don't invent. One structured call fills every field: migrate = lift the existing copy, generate = write new copy in brand voice (and flag it). Won't fabricate prices, reviews or credentials. Then code validates + repairs.

in: scraped content + schema + brand voiceout: filled rows, generated fields flagged

▼

Review human gate run-to-csv.mjs

Rows land in a CSV (the Airtable stand-in). The reviewer only scrutinises the generated fields; the migrated copy is their own words. Edit + sign off.

↑

Content learning — per client

Every copy edit teaches the fill step this client's voice, so the next batch reads more like them. Kept per-client because voice doesn't transfer.

↺ back to step 4

▼

Publish next

Signed-off rows push to Webflow (via Airtable if the client prefers editing there). This last mile is a known API write, deliberately left until the hard part above is proven.

The point: humans only touch the two gates (match sheet + review), and every touch is captured. Structure corrections make matching smarter for all clients; content edits make the copy sound more like this client. The tool compounds.

Content Migration Engine · Greg × Revamp Digital · pilot: Sparkling Pools

Questions on the flow? Email Greg