How we use language models in content editing at 4082
Models now draft alt text, check links and suggest metadata inside the CMS. Editors still decide what gets published. Here is our setup.
Ines Kortmann
Most of the sites we build are edited by small teams: a marketing lead, a part-time writer, sometimes an intern. They do not need a writing robot. They need help with the tasks nobody enjoys.
Three jobs we hand to a model
Alt text. When an editor uploads an image, the CMS suggests a description. The editor accepts, edits or rejects it. Acceptance rate on one client site: 74%.
All three tasks share one rule: the model suggests, the editor decides. Nothing reaches the live site without a person pressing publish.
Metadata. Page titles and descriptions are drafted from the page content and checked for length.
Link checks. Before publishing, a model reads outgoing links and flags pages that moved or no longer match the anchor text.
Jobs we keep human
Headlines, product claims, prices and anything legal. A model can draft these, but in our experience editors spend longer fixing the draft than writing from scratch.
Where the model runs
Calls go through a small service we host per client. Content never trains a vendor model, and every suggestion is logged with the editor’s decision.
If you edit a site we built and want these tools switched on, ask your project lead.
We switch each tool on per site and per editor. Nobody gets a feature they did not ask for.
After a month we asked eight editors on three sites what they would keep. All eight kept link checks. Six kept alt text suggestions. Four kept metadata drafts; the others found it faster to write titles themselves.
What editors told us
Running costs are small. On a site with 1,200 pages and 30 edits a week, model calls cost under €10 a month.
Adding the three tools to an existing site takes about four days of work: one for the service, two for the CMS panels, one for testing with the editors who will use them.
Setup time and cost
One example, start to finish
An editor at a rental company uploads twelve photos of a new apartment. The CMS suggests a description for each. She accepts nine, rewrites three that missed the view from the balcony, and moves on. Before, alt text on that site was empty for 60% of images.
Before she publishes, the link check flags two outgoing links to a city page that moved last month. She updates them in place.
The whole process adds about a minute to a task that used to take twenty, because nobody wrote alt text at all.
Limits we set
Models never see draft content marked confidential. Suggestions expire after a day if nobody acts on them. Editors can switch any tool off for a single document.