Brand DNA

Brand is how Mediastilo learns to write like you. It captures your voice, your typical topics, your structure and formatting habits, and the words you never want to use, then applies all of that every time it drafts an article.

The better your Brand, the more on-brand your drafts. You can set it up in minutes and keep refining it over time.

Your voice profile

The top of the Brand page is the voice profile: eight things Mediastilo tries to learn about how you write.

Each axis shows what we believe, how sure we are, and what taught us — your material, your picks, your edits, or what you published. A long belief is trimmed to fit its card, so no one axis stretches the board; the whole of it is a click away.

Every draft is written from these eight beliefs, so this page is also the honest answer to "why did it write it like that?".

An axis Mediastilo hasn't learned yet doesn't sit empty — it asks. Which sounds like you? opens a pair written for that axis, directly under the board. If you'd rather just say it, Set it is underneath.

Opening one axis

Edit on a card opens that axis on its own, with everything about it in one place:

• The belief itself, editable. Anything you type in is pinned: Mediastilo will not overwrite it with anything it later infers.
• Room for a sentence or two — 400 characters. Each of the eight goes into the prompt for every draft you generate. A long one crowds out the other seven and makes every draft cost a little more, so an axis is a description of how you write, not a brief.
• Ask me about this one. A pair written to settle that axis, rather than waiting for Mediastilo to get to it. It costs one credit like any other pair, and it isn't offered for an axis you've set by hand — your value wins either way, so there would be nothing to learn from your answer.
• How this belief got here. Every time this axis moved: what moved it, what it said before, and what it was reading at the time.
• Forget this, if you'd rather it went back to learning that axis on its own.

What Mediastilo learns from

Any one of these is enough — you don't need all three.

• Your website. If your site publishes a feed, Mediastilo reads a sample of your recent posts. This is the fastest route to an accurate voice, because it's based on work you've already published. A site with no feed is fine, and a website is entirely optional.
• Writing samples. Paste in anything that already sounds the way you want to sound. Remove samples at any time.
• The open web. With no site and no samples, Mediastilo researches the brand by name and works from what's published about it.

It also learns from your edits. When you substantially rewrite a draft — not a typo fix, a real reworking — Mediastilo reads what changed and records what your version demonstrates about the voice. These beliefs are held loosely: they carry the lowest confidence of any source, so a single rewrite nudges rather than decides, and anything you set by hand always wins. It reads at most one edit an hour, and it costs you nothing.

And it learns from what you publish. When a draft goes out, Mediastilo compares what it wrote with what you actually sent, and reads the difference the same way — but files it more firmly than an ordinary edit, because a publication is the version you put your name on rather than whatever you happened to have typed when the editor last saved. Every publication is read; the once-an-hour limit doesn't apply.

Analysis uses credits, because Mediastilo runs an AI analysis each time. See Credits & Billing.

Sharpening the voice

Describing your own writing is hard; recognising it is easy. So instead of an interview, Mediastilo shows you two short samples written to the same brief, which differ in exactly one way — the one thing it is currently least sure about.

Pairs open under the voice board itself rather than in a section of their own: either an axis card that hasn't been learned yet asks for one, or you start a run from the board.

Pick the one that sounds like you and it moves that belief, then goes straight to the next pair. Everything else is held steady, which is what makes a single pick informative. If neither sounds right, say so — Neither sounds like us is a real answer, not a skip, and after two of them Mediastilo stops asking about that axis.

Progress shows as sharpness: how settled the voice is overall, not a question counter. Stop whenever you like and pick it up later; a few picks is usually enough. Each pair costs one credit.

The occasional question in the editor

The same question sometimes appears in the editor, under a draft you've just been rewriting — at most one a day, and only after you've actually changed something. When your rewrite touched a particular part of the voice, that's the part it asks about, so the question follows the work you were already doing.

Close it with the ✕ and nothing is spent and nothing is decided; unlike Neither sounds like us, it doesn't count against that axis, and the question can come back another day.

What it wrote

Under the voice profile, What it wrote shows the most recent draft that was written from your current beliefs, with an excerpt and the list of traits that were in force when it was written. It is the shortest answer to "is any of this actually doing anything?" — and if a piece reads wrong, it tells you which beliefs to look at. Open the draft to read the whole thing.

What changed

What changed is the record of every movement in your voice profile, newest first. Each entry says which axis moved, what it moved from and to, what taught it — your material, a pick of yours, one of your edits, or something you published — and, where there is one, the passage that was the evidence.

Three things can happen to an axis, and the entry says which: it can be learned for the first time, changed from one belief to another, or forgotten.

Anything Mediastilo worked out from your edits, without you being asked, is marked needs a look and sits at the top until you deal with it. Two answers:

• Keep — you agree; the entry stops asking.
• That's not right — the belief is undone. The previous one comes back exactly as it was, or the axis goes back to empty if there was nothing there before. The entry stays in the list, marked as undone, so the record is honest about what was tried.

Everything you actively chose — a pick in a voice comparison, anything you typed in — is recorded too, but is not flagged: you already know about it.

How the voice profile reaches a draft

Every belief in the profile is sent to the writer when a draft is generated, and each one is stated with the strength we actually hold it at. A value you set yourself arrives as an instruction; something we are confident about arrives as a rule; something we only lean towards arrives as a lean. That is why a half-formed profile is still worth having — it steers without pretending to be certain.

Where a belief here conflicts with what we learned by reading your existing site, the profile wins. The site analysis describes what you published before working with us; the profile describes what you have asked for since.

What you cover

The voice profile is how you write. What you cover is what you write about, and it is counted rather than inferred: no model reads anything here, so it costs nothing and can always show its working.

Three things count, because each is an unambiguous act by a person:

• Publishing a draft. Every topic in the material behind it gets credit.
• Turning a draft down as Off our beat. The exact opposite of publishing it, and weighted just as heavily — both are judgements you made after reading the finished writing.
• Reaching into a run to Include an item, or to throw one out. These count for less: they're a call made on a headline, not on a finished piece.

What the engine decided on its own is deliberately not counted. A scorer fed its own past decisions would only learn its own bias back and call it your preference.

Each entry lands in one of two columns — More of this or Less of this — with the count that put it there ("2 published", "1 turned down as off-topic", "2 kept by hand", "2 thrown out"). Nothing appears until two deliberate acts point the same way: one force-include is a Tuesday, two is a pattern.

These preferences steer the scoring in your next run — a topic you cover is weighted up, a topic you keep throwing out is weighted down. Nothing already scored changes.

A new entry is marked New and waits for you:

• Keep — you agree; it stops asking.
• That's not right — it stops steering immediately. It won't come back on your next click either; only materially more evidence than the evidence you already rejected can raise it again.

An entry also disappears on its own if the evidence stops supporting it — undo a force-include, and the count falls back under the line.

The number next to Brand

The count beside Brand in the sidebar is those unreviewed beliefs, plus any topic preference you haven't looked at, plus any brand suggestions you haven't accepted or dismissed. All are finite, so the number reaches zero and stays there until something genuinely new happens. It is deliberately not a count of available questions — sharpening never runs out, and a badge that can never be cleared is one you'd learn to ignore.

Opening Brand shows the same number broken down at the top of the page — Voice settings to review, Topic changes to confirm, Advice to read — and each one jumps to the section holding it. The banner is only there when something is genuinely waiting, so if the page opens without it, the number was zero and there is nothing to do.

Brand suggestions

After analysis, Mediastilo proposes suggestions to strengthen your brand, each one is an observation plus a concrete recommendation, grouped into:

• SEO: discoverability improvements.
• Style: voice and tone refinements.
• Consistency: keeping your output uniform.
• Engagement: making content more compelling.

Review them and accept the ones you agree with; accepted suggestions feed back into how Mediastilo writes for you.

Custom instructions and prohibited words

• Custom instructions: free-form guidance Mediastilo should always follow when writing for you (available on paid plans).
• Prohibited words: terms Mediastilo should never use. Add anything off-limits for your brand and it will steer clear.

Structure: the shape of a piece

Voice and structure are two different things, and Mediastilo learns them in two different ways. Your voice is inferred from what you've published. Your structure is declared by you, in the Post structure section of the Brand page.

A content format is a reusable skeleton for a piece:

• Sections, in order, each with a heading and an instruction describing what belongs in it ("Good to know, practical detail the reader needs to act, dates, prices, how to get there"). Mark a section required and it is never dropped.
• Key facts, the labelled basics a reader wants up front, "Where", "When", "Price", "Who it's for". They render as a short block at the top (or bottom) of the draft.
• Options: heading level, whether to use bullets, a target length, and where the key facts sit.

Without a format, Mediastilo still writes with headings and concrete detail. With one, it writes to your shape, every time, including the drafts it generates overnight without you watching.

Building a format from examples

If you already have posts shaped the way you want, you don't have to describe the shape by hand. Paste two or three of them into Derive from examples and Mediastilo reverse-engineers the skeleton, its sections, its key-fact fields, its formatting, and nothing about its subject matter. The result is a draft format you review and edit before saving; it isn't saved for you.

The default format

Mark one format as default and every draft for that brand uses it unless the draft names another, including the drafts the overnight pipeline writes for you. Formats belong to a brand, so each brand in a workspace can have its own.

Missing facts are asked for, never invented

When your format asks for a fact the source material doesn't contain, Mediastilo does not make one up. It leaves the fact out and records it as a gap on the draft, a short note like "Ticket price not stated in the source". You'll see the gaps on the draft in the editor. They are the questions the piece needs answered before it's publishable, and answering them is your job, not the AI's.

Refining over time

Your voice evolves, and so should your Brand. From the Brand section of your dashboard you can re-analyze, add or remove samples, accept new suggestions, and edit your instructions and prohibited words whenever you like. Refining your brand also uses credits.

Where it's used

Your Brand powers the whole pipeline: it scores how well incoming source material aligns with your brand (in Signals) and shapes every draft Mediastilo generates (in Drafts). See Sources and Drafts & Publishing.