Rewriting articles
Purple Prompts

From the dropdown (1), you can choose one of the prompts defined in Prompt Management, and see the details of this prompt (2).
You can decide how you want the rewriting to work:
- Should it delete the old text and replace it with the new one (3), or
- should it instead append (4) the newly generated text below the existing article text?
With a click on 'Rewrite article' (5), a popup opens, and the language model starts working.

Alternatively, you can select 'Custom...' from the dropdown (1). This enables you to edit a prompt spontaneously right here by typing your prompt into the text field (2).
The model parameters, the temperature (3) and max tokens (4) can also be adjusted.
The Replace (5)/Append toggle (6) and the Rewrite button (7) work the same as with a predefined prompt (see paragraph 'Using predefined prompts' above).

Once started, the rewriting process gets shown in a popup dialogue. You can see the original post content in text form (1) on the left (this is what the language model works with), and you can watch the AI output being generated (2) on the right side.
Within this dialogue, we try to detect important phrases and colour-code them, to give you content confidence at a glance:

Green phrases are detected in the left and right text (e.g. 1, 2, 3 in the screenshot above) and indicate that this information was preserved by the language model.

Grey phrases are detected on the left, but not on the right: The language model might have ignored or omitted information here. E.g. 1 in the screenshot above. In this example, the language model has only taken part of the phrase. As the word 'ursprünglich' wasn't taken into account, the phrase was marked as 'grey' on the left side, but green on the right one. For summarization tasks, this can be desired.

Orange phrases are detected on the right, but not on the left: The language model generated this text, but it was not found in the original text. This might be due to re-phrasing or synonym usage, but it might also indicate cases of hallucinations.
Please review Orange phrases closely. Should the model actually imagine wrong information here, see also if you can adjust your prompt as described in Prompt Writing Best Practices.

After reviewing the generated text, you can either click 'Accept' (1) to take it like it is, or click 'Generate again' (2) to generate a new variant of the text.
The 'Generate again' button restarts the language model and produces a new output. Depending on the prompt text and also especially the temperature setting, the newly generated text might be exactly or nearly the same as before, or completely different.
See Prompt Writing Best Practices for more information.

It is possible to rewrite only parts of a post, rather than rewriting the whole post at once. To do this, simply select text in the block editor. Only the selected blocks (1) will be sent to the language model. The current mode, whether the model should work with the whole article or only with some blocks, is also communicated via the sidebar panel (2).