Writing style

Writing style lives on the Voice tab of Content settings, at Content → Settings → Voice. You give it finished writing you are happy with, and it builds a voice profile that every draft is written against.

Do this once per brand, early. Without a profile, drafts still avoid the worst AI tics, but they land on a generic house style rather than yours. With one, the model is told your formality, your sentence shape, the phrases you reach for, and the phrases you never use, as a hard drafting constraint rather than a suggestion.

Voice is separate from the knowledge base. Knowledge decides what a draft can say. Voice decides how it says it.

Adding samples

The Samples section takes three kinds of input, and the header asks for three to eight finished pieces.

• File accepts PDF, DOCX, TXT, and Markdown, up to 5MB. Upload anything that looks like code and Trakkr asks you to confirm first.
• URL scrapes the page and keeps the extracted article text.
• Paste text takes a name and a body, for drafts, memos, and anything unpublished.

Each sample must be at least 100 words, or it is refused. Identical content is refused too, because samples are hashed per brand. Anything over roughly 15,000 tokens (about 11,000 words) is accepted but truncated, so a book chapter contributes its opening, not all of it.

Every row shows the source, the word count, and the status, and each has a delete control.

What the analysis reads

The analyzer is asked for five things, and the profile it returns is free-form JSON rather than a fixed set of fields:

• Voice and tone: overall characterization, formality, confidence, relationship to the reader, emotional register.
• Structure patterns: paragraph length, use of lists, how you open a piece, how you transition, how you close.
• Sentence style: typical length, fragments, rhetorical questions, active against passive, asides.
• Vocabulary: technical depth, industry terms you use naturally, signature phrases, constructions you avoid.
• Distinctive elements: what makes the writing recognizable, quirks, how you handle examples and analogies.

If your samples disagree with each other, the analyzer is told to say so and describe the range rather than average them into mush.

Reading the voice profile

Under Voice profile the tab reports one of four states.

Once it is ready, View profile expands up to ten labelled rows drawn from the analysis: overall voice, formality, register, reader stance, sentence length, pacing, vocabulary, preferred phrases, avoided phrases, signature devices, and so on. The labels vary with what the analyzer found, because the profile is not a fixed schema.

When the analysis runs

You do not have to trigger it. Adding or deleting a sample queues a fresh analysis in the background. Before it spends anything, Trakkr hashes the combined text of every ready sample and compares it to the last analyzed hash, so nothing re-runs when nothing changed.

Re-analyze in the tab header forces a pass anyway. It is disabled while you have no samples. Delete every sample and the profile is cleared rather than left behind.

If a run fails, the profile records the error and the tab reports it. The most common cause is a change of plan or brand access mid-run: analysis re-checks paid editor access before it calls the model and again before it saves, and cancels rather than storing a result the account is no longer entitled to.

How the voice reaches a draft

At generation time the profile is folded into the drafting instructions as a constraint, not a hint. Preferred vocabulary and signature phrases are passed through as words to reach for. Anything the analysis recorded under avoided phrases is passed through as words to keep out. The formality reading also picks a pair of calibration examples, one good and one bad, so the model has a concrete target rather than an adjective.

Two things happen with or without a profile. Drafts get a standing list of AI cliches to avoid, and they are told to open sections with a specific answer or tradeoff rather than trend language. A voice profile sharpens that floor; it does not create it.

Word-level guardrails are separate and manual. Content → Settings → Identity has Preferred and Avoid phrase lists that apply regardless of what the analysis found, which is the right place for a term legal insists on.

Common questions

How many samples should I add?

Three to eight, and the hard cap is ten per brand. Below three the analyzer struggles to tell voice apart from subject matter. Above eight you are mostly adding confirmation, and only the most recent ten are used anyway.

Why does the draft still not sound like me?

Usually the samples. Check that they are genuinely yours rather than ghostwritten or heavily edited by someone else, that they are recent, and that they agree with each other. Voice transfer also has a ceiling: expect a draft that fits alongside your work after an edit, not a draft indistinguishable from it.

Does the profile apply to campaigns as well as single articles?

Yes. The profile is loaded per brand whenever an article is assembled, so every draft made from that brand uses it, however it was started.

Can I edit the profile by hand?

No. The profile is written by the analysis and there is no editor for it. To change it, change the samples and re-analyze. For a specific word you must include or exclude, use the phrase lists on the Identity tab instead.

What happens to samples over the length limit?

They are truncated to roughly 11,000 words on the way in, and the combined set is capped again at analysis time. Very long documents therefore contribute their opening. Trim to the section that best represents your voice rather than relying on the cut.

Do writing samples feed anything besides voice?

No. Samples are read for style only. Facts in them are not quoted or retrieved. Anything you want a draft to cite belongs in the knowledge base.