Discourse markers in writing

By Jack Stovell · published 2026-09-24

Discourse markers are the connecting phrases you reach for when you move from one thought to the next, such as "so", "here's the thing" or "and to be fair". A ScriptGrain voice profile stores up to five, quoted verbatim from your samples, and they feed the signature-phrase check in the voice match score at weight 2.

Analyse your discourse markers free

Your free voice profile measures discourse markers among all 45 attributes, from a sample of your own writing. No card needed.

Try the six-signal check first

At a glance

What discourse markers measures

This attribute measures the specific words you use to link one thought to the next. It sits in the Function words group alongside four other attributes, and the app labels it "Discourse Markers" with the API key discourse_markers. The scale is a list, capped at five phrases per analysis, stored as a string array in the profile.

What sets this attribute apart from most others is that it has no scale of high or low: it catalogues what you actually say. If you write "at the end of the day" three times across your samples, that phrase can earn its place on the list.

The extraction model (Claude Sonnet 5) builds this in two passes: pass 1 reads each sample and pulls candidates, pass 2 synthesises across all samples into a final list. Nothing here is invented or generalised from a corpus. It's collected verbatim, which is why the phrases in your profile sound like you and not like a style guide.

How discourse markers is scored

Discourse markers move the voice match score through a shared feature called signature phrases, weighted 2: the pool combines your preferred_words, discourse_markers, filler_phrases and preferred_phrasings, keeping multi-word phrases and single words of six or more letters, up to fifteen entries total. The feature needs a pool of at least three entries to count, and a one-word marker such as "so" is too short to enter it.

Similarity is worked out as hits divided by the smaller of 4 and the pool size, then capped at 1. So if your pool has three phrases and a draft uses two of them, similarity is 2 divided by 3, about 0.67; four hits from a pool of four or more scores the full 1. Code counts the hits in the draft; no model judges this step, although the list itself is the extraction model's reading of your samples. Full mechanics live on the voice measurement framework page.

There's no reference corpus for this attribute, so there's no "typical" phrase list to compare against. It is one of 28 of the 45 attributes that can move the score.

How to analyse discourse markers in your own writing

You don't need software to spot your own habits here. A read-through and a bit of patience will do it.

  1. Pull three or four pieces you've written recently, ideally different formats: an email, a blog post, a social caption.
  2. Underline every phrase that bridges two sentences or two ideas: "so", "and to be fair", "here's the thing", "except".
  3. List the phrases separately and count repeats. A phrase you use once might be accident. A phrase you use five times is a habit.
  4. Cut it down to your top three to five. That's your actual discourse marker set, the one a reader would recognise as you.
  5. Check whether any of them are stock connectors you picked up from formal writing training rather than from how you actually talk. Those are worth dropping.

ScriptGrain's free voice profile reads this attribute from a writing sample too, alongside the other 44 attributes.

Examples of discourse markers in real writing

Example: "So, here's what actually happened." The word "so" does the work of a full transition sentence in one syllable.

Example: "And to be fair, the numbers back that up." Two markers stacked together, softening a claim while still making it.

Example: "That's why the second draft always reads better." A causal link stated flatly, no hedge, no build-up needed.

These are illustrations only; there's no reference set for this attribute. They're just how the link between two ideas tends to sound out loud.

How to make AI use your connecting phrases

Getting an AI assistant to sound like you starts with telling it exactly which phrases are yours.

  1. List the connecting phrases you actually use, the ones from your own analysis above, not a generic style guide.
  2. Give the assistant one or two examples of each phrase in a real sentence, so it sees the phrase doing a job, not sitting in isolation.
  3. Tell it to place them where you naturally would, at a pivot in the argument, not at fixed intervals like every third paragraph.
  4. Allow, don't mandate. A marker forced into every section reads worse than one left out.

When a ScriptGrain profile generates a draft, this list is sent inside the voice profile JSON block of the system prompt, under an instruction to treat every attribute as a hard constraint. ScriptGrain's generator also lifts its bans on stock AI phrases for any phrase in this list, so a marker you really use stays allowed. The profile reaches ChatGPT, Claude and other tools through the API and MCP server, so the same list travels wherever you're actually writing.

How to make AI use fewer stock connectors

Banning specific words works better than asking for "natural" writing in general.

  1. Name the stock connectors you want gone; "Moreover" and "Furthermore" are the usual offenders.
  2. Tell the assistant it's fine to let two sentences sit next to each other with no marker at all, when the link between them is obvious anyway.
  3. Once you have a draft, scan it for one word repeated across several paragraphs. That's usually the tell that a model has fallen back on a habit rather than following your voice.
  4. Swap the repeat for one of your own phrases, or cut it entirely.

A ScriptGrain draft works from the other side: its generation prompt bans a list of stock AI phrases and lifts a ban only for a phrase that appears in your profile's word lists, this one included. Your contrast word and paragraph openers are worth checking in the same pass.

Questions

How are discourse markers measured?

They're collected verbatim by the extraction model across two passes: one per sample, one synthesising across all samples. The result is a list of your own phrases, capped at five per analysis. There's no scale of good or bad here, no reference corpus comparison; it's a catalogue, though it feeds the signature-phrase check in voice match scoring at a weight of 2.

What is a discourse markers list?

It's the set of connecting phrases a writer actually uses to link thoughts, things like "so", "here's the thing", "at the end of the day". In a ScriptGrain profile it's a string array of up to five entries, quoted directly from writing samples rather than drawn from a generic list of transition words.

Is there a linking phrases checker?

Not a standalone tool, but the analysis works by hand: read your own writing, underline every bridging phrase, count the repeats. ScriptGrain's free voice profile does this automatically as one of 45 attributes it extracts from a sample, alongside your contrast word and paragraph openers.

How do I make AI use natural transitions?

Give it your own list of connecting phrases with examples in context, then tell it to place them where the argument actually pivots, not on a fixed schedule. That's the difference between transitions that read like you and transitions that read like a template.

How do I make ChatGPT stop saying "moreover"?

Ban the word directly, and give the assistant permission to leave sentences unmarked when the connection is obvious. Also feed it your own preferred phrases as replacements; an outright ban with nothing to replace it tends to just produce a different generic connector instead.

Related

Sources