Revising and rewriting AI texts

Revising texts with AI in one sentence

Revising texts with AI means improving an existing text very specifically by giving a separate and narrow prompt for each task. This works much better than asking for a general improvement, because this is exactly where the difference between a usable and an annoying revision lies.

An open prompt almost always leads to the model rewriting the text completely, so that important statements often disappear or numbers change. A narrow prompt, on the other hand, changes exactly what you want to change, so that your own voice is preserved.

The five types of revision

Each type of revision is allowed to change something else, which is why each task gets its own pass with clear rules.

Never mix two types in one pass, because otherwise you will get a text whose deviations can no longer be controlled. If you shorten and change the tone one after the other, it only takes thirty seconds longer and is significantly safer.

The last column shows how you determine how strong the intervention should be. Some writing programs have a slider from "light" to "strong" for this. In a chat with a language model, this slider does not exist, so you write the limit as a concrete rule in the prompt, for example "only touch sentences over 25 words".

One paragraph, revised five times

A concrete example shows best how differently the results turn out when you apply different types of revision. The following original paragraph comes from an older text about monthly reporting. Each of the five versions starts fresh from the original, so that you can see exactly that each type of revision touches something else.

Original, 67 words

In today's digital world it is important to note that a monthly reporting that brings together the relevant key figures from different sources can help to potentially improve decisions in marketing. We create reportings that are faster, cheaper and more efficient, and that seperately for each channel. This makes it clear that reporting is an important part of every marketing strategy, which is why we created over 200 reportings last year.

1. Corrected

In today's digital world, it is important to note that a monthly reporting, which brings together the relevant key figures from different sources, can help to potentially improve decisions in marketing. We create reportings that are faster, cheaper and more efficient, and that separately for each channel. This makes it clear that reporting is an important part of every marketing strategy, which is why we created over 200 reportings last year.

Three missing commas were added and a typo was corrected, but otherwise not a single word was changed. The paragraph is now error free and still bad in terms of content.

2. Edited

A monthly reporting brings together key figures from several sources and thus improves marketing decisions. We create these reportings separately for each channel, faster and cheaper than a manual evaluation. Last year there were over 200.

Nested sentences are resolved and vague phrases are replaced by clear statements, so that the text becomes more understandable. Empty phrases were deleted and the rhythm reduced to provable points, while the core statements and the number are preserved.

3. Rewritten

Monthly reportings bundle numbers from different systems in an overview, on the basis of which marketing teams make better decisions. Every channel gets its own evaluation with us, and that is faster and cheaper than manual work. 200 such evaluations were created in the past year.

Word choice and sentence structure are completely new, the order remains untouched. No statement was added or removed.

4. Shortened, target number 25 words

Monthly reportings bundle key figures from several sources, separated per channel. This accelerates decisions and costs less than manual work. Over 200 pieces last year.

Secondary information is deleted, because the requirement was a fixed target number and not a general shortening. The core statements and the number 200 were marked as untouchable in advance, so that they safely remain in the text.

5. Tone changed, simple language

A reporting is a monthly overview. It contains the most important numbers from all channels, separated for each channel. This way you quickly see what works. Last year we created more than 200 of these overviews.

Technical terms are explained and the sentences are limited to a maximum of twelve words, so that the text is easier to read. A direct address was introduced, while content and number remain identical.

In your own work it helps a lot to compare before and after, instead of reading both versions completely. If you have the readability measured, it ends any discussion about whether a text is really phrased simply enough.

[[IMAGE: Screenshot from our own diff tool, original and edited version side by side with marked changes]]

The editing prompt: what belongs in it

An effective revision prompt always consists of four parts, because the model needs exact specifications. You need a clear task, fixed boundaries, a target picture and an output format.

Edit the following text for comprehensibility. Do not change any statements, any numbers, any order of the paragraphs and any technical terms. Resolve nested sentences, remove word repetitions, improve transitions. Target audience are marketing managers without a technical background. Output the edited text, below it a list of the most important changes.

Three details make the difference here if you want to achieve good results. You must explicitly name what is forbidden, because clear prohibitions work more reliably than mere hope. Always request a list of changes, so that you keep an overview, and work section by section, because the revision often deviates from the goal in long texts.

The strongest lever for a good style is a reference text, which the model can use for orientation. This prompt can be copied directly, so that you can use it immediately.

Here are three paragraphs from our existing texts as a style template. Analyze sentence length, address, technical depth and promotional tone of this template invisibly, but apply them. Rewrite the following text in the style of the template. Keep all statements and numbers. Do not use any of these words: [Negative list].

How prompts are generally structured is detailed in the article Writing better prompts. The most important part of an editing prompt always remains the list of things that must remain unchanged, so that no important information is lost.

Spelling: why the language model is not ideal here

A language model corrects spelling very well. However, it poorly adheres to the rule of only fixing errors. Almost every model also smooths out the style unprompted and deletes word repetitions, which is a real problem with an approved text. Not a single word may be changed in such texts.

Two antidotes help here, so that you keep control. Either use a tool that only marks errors, or request an error list with location, error and suggestion, so that you can implement this list yourself. The second way is the only safe one for legally checked texts, because this way you avoid unwanted changes.

Which tool for what

There is a suitable tool for every task. The choice determines how much control you keep over the text.

The most important row is the second one, because for coordinated texts a tool with change tracking and block list beats the free prompt. This is not due to the intelligence of the tool, but because you see every deviation immediately. Which provider belongs in which category changes constantly, which is why you should decide based on the tool type. Check carefully whether the change tracking is included in the chosen plan.

Controlling style and tone specifically

The request for a more professional text is not an instruction, but only a wish that often leads to inaccurate results. However, you can reliably control the tone via three levers, so that the text sounds exactly as you want it to.

1. Description in dimensions. Name several axes instead of a single adjective, by using specifications like sentence length, direct address, technical depth and promotional tone, because this can be reproduced well.
2. Examples instead of description. Provide two to three paragraphs from your texts and have it write in the style of these templates, since models imitate concrete role models much better than abstract descriptions.
3. Negative list. Define words and constructions that may never occur, because such a list is more valuable in the long run than any style description.

Our own negative list begins with the following entries, which you are welcome to adopt and expand for yourself.

• Filler words for transitions, such as furthermore or not least
• Empty phrases like tailor-made or innovative
• Imagery from advertising copy, for example to the next level
• Time clichés like in times of
• Amplifiers without proof, such as revolutionary or unique
• Dashes as punctuation marks, because period and comma are completely sufficient

There are two important rules for this that you should keep in mind. Do not forbid any word generally, but only its empty use, so that the text remains natural. Also maintain the list from real corrections and not from theoretical guides, because that is much more practical.

Recognizing and removing AI writing patterns

Whether a text comes from a machine or from a human who writes like a machine is completely irrelevant to readers. The typical patterns are annoying in both cases, which is why these six points are worthwhile as a fixed checklist for every text.

• Rule of three. Phrasings like faster, cheaper and more efficient seem charming once, but tiring four times.
• Announcement sentences. Sentences like In the following we will look at can be deleted without replacement and you can start directly.
• Softener chains. Constructions like can help to potentially improve should be replaced by a clear statement.
• Uniform paragraphs. If every paragraph has exactly four sentences, it looks unnatural. Real texts have short and long passages.
• Summarizing concluding sentences. A conclusion in every section is too much. That belongs at the end of the text at most.
• Overemphasis of the obvious. Phrases like It is important to note only bloat the text.

In practice, this works as a separate pass, where you give the text plus checklist to the model. You demand to exclusively remove these patterns. Afterwards, a human reads over it. What else indicates machine origin is explained in the article Detect AI content.

Sentence rhythm as an instruction, not as a wish

Uniform paragraphs do not disappear by asking to vary the sentence rhythm, because an adjective like "varied" is not a clear specification for a model. A number is the clearer specification, even if the model only roughly adheres to it.

Imagine a recipe. If it only says "sweet", nobody knows how much sugar is meant. If it says one hundred grams, there is a direction that everyone can orient themselves by. A number in the prompt works similarly, it replaces the feeling with a goal. This prompt has proven itself.

Revise the sentence rhythm of this text. At least one sentence with a maximum of six words per paragraph. No sentence longer than 28 words. No two consecutive sentences with similar length. Do not change any statements and any numbers.

Before: A monthly reporting bundles key figures from several sources and thus provides the basis for decisions that would otherwise have to be made from the gut. After: A monthly reporting bundles key figures from several sources, which replaces the gut feeling. You check yourself at the end whether the specifications are really met, for example with a quick word count, because language models do not count reliably.

What humanizer tools do not achieve

Humanizers, i.e. tools for disguising AI texts, often promise that AI texts are no longer recognizable as such. However, this promise does not stand up to close scrutiny.

First, a detector score is not a real quality standard, because detectors make mistakes in both directions. Human technical texts are often classified as machine-made, while revised AI texts remain undetected. Second, unrecognizability is not a meaningful goal, since search engines evaluate based on the benefit for the reader and not on the origin.

Third, a humanizer does nothing that a narrow editing prompt with a good checklist cannot also achieve. In return, it costs you the control over every single change. We measured this ourselves on 100 machine and 100 human texts each, as our investigation on AI text detectors shows.

Rewriting without plagiarism

When rewriting third-party texts, there is a very clear legal boundary that you should know. A text that has merely been swapped word for word remains a modification of the original under copyright law. Only when the new work maintains a sufficient distance from the used text is it no longer a modification, because the structure, the train of thought and the arrangement of the material can also be protected. The adoption of pure facts and information is permissible, but not that of the individual presentation.

You work cleanly by not having the third-party text rewritten, but instead extracting the core statements. Formulate a new text from this with its own structure, its own weighting and its own examples. This is also the better way in terms of content, because a merely swapped third-party text offers no new value. For literal adoptions, the right to quote with exact source citation still applies.

This can be checked with a plagiarism check, which finds literal adoptions. You must know its limit. It does not recognize adopted structure and train of thought, and exactly these can be protected.

In which order you revise

The five types of revision have a logical sequence, which is often different than one initially expects.

1. Shortening. Decide first what actually stays in the text, because any work on text that is later deleted is wasted time.
2. Editing. Improve the comprehensibility and the sentence structure on the already reduced text.
3. Tone. Only when the content is fixed is it worth adjusting the address.
4. Removing writing patterns. Do this as a separate pass with a fixed checklist, so that the text looks natural.
5. Correcting. Do this at the very end, because every previous pass can introduce new errors. Whoever corrects first edits sentences that might not even exist at the end.

Rewriting stands outside this chain, as it is not a single step. It is rather a special case when you need the same content in a completely new phrasing.

Conclusion

Revising is the discipline in which AI unfolds its most reliable benefit. However, this only applies if each pass has exactly one task and clear boundaries, so that you keep control.

Use narrow prompts, request a list of changes and control the style via examples instead of adjectives. Work through writing patterns as a checklist and always correct at the very end. The last look remains human, because it only takes a few minutes and decisively determines the impact of your text.