How to build a house style from phrases you never use

Most style advice is a list of things not to do. George Orwell closed “Politics and the English Language,” printed in Horizon in April 1946, with six rules, and the third is a ban: “If it is possible to cut a word out, always cut it out.” A banned-phrase list is that habit made concrete, and it is the part of a house style a writer can use without rereading the whole guide.

Why is a banned-phrase list the most useful part of a house style?

A banned-phrase list can be checked by searching, and a principle cannot. “Be clear” needs a judgement on every sentence, while “utilise” needs a find-and-replace. GOV.UK’s A to Z style guide works this way: it opens its word list with “Plain English is mandatory for all of GOV.UK so avoid using these words,” and gives each word a replacement. Utilise becomes “use,” and “going forward” becomes “from now on.”

The replacement is half of the entry. Digital.gov’s plain-language guide shows the pattern with a hidden verb: it turns the noun phrase “conduct an analysis” back into the verb “analyze.” An entry without a replacement only tells the writer what failed, and the next sentence fails in a new way.

How do you build the list from your own texts?

A useful banned-phrase list comes from evidence, not from a template. Two sources are worth reading, your own writing and an assistant’s draft on a subject you know.

• Take an assistant’s draft on a topic you know well, and mark every phrase you would not say aloud to a colleague.
• Take two or three pieces of your own writing that you are proud of, and note the phrases that never appear in them. Your own habits belong on the list too.
• Write each phrase exactly as it appeared, then write what you would say instead. If you cannot, the entry is a vague dislike, so drop it.
• Keep the entries that recur across drafts and delete the one-offs.
• Sort what is left by shape: stock openers, hedges, filler phrases, clichés. A pattern is easier to remember than a long run of separate phrases.

The shapes are the ones in GenPolish’s published taxonomy of nine tells. Wikipedia’s editors keep a similar field guide, with example words such as “nestled” and “in the heart of.” That guide cautions that the same small set of promotional phrases turns up whatever the topic, and that not all promotional writing is AI-generated. Treat your list as a rule about your own prose, not as a test of anyone else’s.

What does a starter list look like?

Copy the table, delete what you would never be tempted by, and add the phrases your own review turns up. The last row is the one that matters, because a list built from your drafts fits your voice better than any published list.

How do you use the list in an assistant’s instructions?

Assistants read a standing set of instructions before each reply, so the list belongs there. Where that setting lives differs by product and changes over time, so look in your assistant’s settings for custom instructions or a style option. Write each entry as an instruction to do something. Anthropic’s prompting guidance puts the principle in one line: “Tell Claude what to do instead of what not to do”. So write “say use, not utilise” rather than a bare list of forbidden words.

Give the instruction an example of your own sentences as well. The earlier article on making ChatGPT sound like you wrote it covers that step. A list of replacements describes what to avoid, and a paragraph of your writing shows what to aim for.

How do you use the list when you review a draft?

The list earns its place in review, because an assistant will slip. Search the draft for each entry, read every hit in context, and rewrite it with the listed replacement. A word-for-word swap sometimes leaves a hollow sentence, so ask what the sentence was meant to say.

GenPolish’s humanizer does the same pass on a pasted text and shows every change as a diff, so you can accept the replacements you agree with and reject the rest. The text-diff checker compares two versions of a draft, and the em dash cleaner handles one punctuation habit with its own page. None of this proves who wrote a text, and cleaning a draft of banned phrases is not a way to pass a detector.

Why does the list need updating?

A banned-phrase list ages, because the habits it targets move. Wikipedia’s guide notes that older models such as GPT-4 wrote more blatantly positive text, while newer ones are more subtly positive and avoid superlatives like “the best.” Reread an assistant’s latest drafts every few months, retire entries that no longer appear, and add the new ones with their replacements.

A banned-phrase list catches phrasing you can see. Anthropic says Claude’s text watermark applies only to words the model chooses and works by changing how it picks among them, so no list of phrases reaches it, and light editing probably will not remove it completely. A complete rewrite changes those choices. That is what GenPolish’s Strong setting does, and it makes no promise about any detector’s verdict.