Text-to-Template

A relaunch offers content momentum: large amounts of content need to be revised. To ensure high-quality results even in decentralized teams, Liip developed TextMate.

TextMate standardizes texts across the following dimensions:

• Spelling & inclusion
• Simplification
• Content guidelines
• Tone of voice

We built TextMate as a typical MVP: few features, thoughtful and effective.

For further development, we’re working on a more complex use case:
We don’t just revise existing content linguistically but have an LLM adapt text to specific page types.

Initial tests are positive. It’s important, however, to have 2–3 good text examples per page type written by a human. Based on these, the system can generate an initial draft that editors can refine.

The advantage: editors don’t start with a blank page—they begin with a first version they can improve.

Are you interested in this approach? Do you have your own use case you’d like to try it on?
Then feel free to get in touch with us!

The Details

How are we developing our MVP?

Anyone who works with an MVP knows this: as soon as the product is in use, the first requests for additional features arise. The key question for the product owner is: in which direction do we develop further? Where are our resources most effectively invested?

One of the most common requests we’ve heard for TextMate: automatically shorten texts.

Sounds simple—but it’s not. Because:

• How much should the text be shortened? There’s no fixed answer.
• Which variations should be offered? And how many?

The more flexibility we allow, the closer we get to free prompting. At a certain point, it makes more sense for editors to work directly with an AI chatbot like ChatGPT.

Another request, however, caught our attention—because it could save humans a lot of repetitive work:

After a test run with TextMate, Clemens Nef, Deputy Head of Communications of the Canton of St. Gallen, gave us detailed feedback. For him and his team, linguistic revision alone isn’t enough. They want structural optimization of content for their website.

Questions like:

• How should the text be structured?
• Which subheadings are needed?
• Where do lists make sense?
• Etc.

A legitimate request. Structure is so important because users scan content on the web—they rarely read linearly. Structure helps users find their way quickly. Search engines and LLMs also prefer well-structured content.

Structure Follows Page Type

But implementing this isn’t trivial either. Structuring isn’t a standardized process.

Websites that optimize user experience work with fixed page types—for example, overview pages, product pages, or team pages. The specific page type defines the structure for the content.

Working with page types is essential—especially for companies with many products, services, or topics. Through page types, we achieve consistency—a key to user satisfaction.

Content written for a page type must be consistent across multiple examples. Design is only the beginning: developing consistent textual and linguistic patterns is the next step—and it’s labor-intensive.

From Design to Text

At Liip, the workflow looks like this:

• Design and content teams jointly develop the page types.
• UX writers create the first examples—ideally based on existing content (e.g., from the old website).
• Insights from this work help iteratively refine the design.
• Based on the finalized page types, all further content is created—consistent in structure, content, and tone.

With each example written, we refine text and language patterns until they’re final. Then they must be applied consistently to all further examples.

In this process, humans do exactly what Clemens Nef wants TextMate to do:

• We optimize page structure and linguistic patterns.
• And apply them consistently across many pieces of content.

Once these patterns are defined, the task becomes repetitive—time-consuming (and therefore expensive) if done by UX writers. That makes it truly interesting to hand over to AI—or, as we say: safe enough to try.

Text-to-Template

We’re now working on automating part of this process through a prompt set:

• The page type is still developed by design and content teams.
• Then we translate the page type into a prompt set.
• Our UX writers create several high-quality examples per page type, which are integrated into the prompt set.
• We feed the system with existing content as a source—e.g., from the old website.
• The system generates new text while applying the patterns of the chosen page type.
• The result serves as a draft that editors can refine.

Initial Insights

Early tests are positive—it seems to work.

Here’s what we’ve learned so far:

• The examples written by UX writers are crucially important.
• The quality of the results also depends heavily on the amount of source content. If there’s very little text on the old website, the system naturally struggles—or fails—to generate high-quality, more detailed results.
• Process optimization will be needed: a key value of the relaunch’s content momentum is that old content is questioned and improved. The system, as currently designed, only does this to a limited extent. Can we explicitly assign this task to editors in the process?
• It’s likely that the description of the page type itself could also be created by an LLM.
• For our next steps, we’re particularly interested in connecting this to the component library in Figma.