seomantik-content

It turns one page of your plan into a finished article, with a record of where every sourced claim came from.

It works from the plan seomantik-topic built: the page's questions, what your site already says, what the pages ranking for the subject cover, and how your site sounds. It writes in your site's own language and voice, checks every sourced claim against its origin, removes the machine tone, and hands you a file ready to publish.

What to expect

Starting

In the project folder where your plan lives, type one of these:

Build the next 3 pages

Build wave 2

Build the page for "the page title or its address from the plan"

"The next 3 pages" takes them in the plan's own order. "Wave 2" builds every page in that wave, one after another.

The questions it asks you

Only at the start of a batch:

• How many pages, to confirm the count.
• Whether your site sells what it writes about (bookings, products, a service), if the plan does not already record it. This decides who your articles may name.

Then it works through the pages without stopping. After each one it reports the page, how many claims are sourced (and says so plainly when there are few), anything it could not source, and the internal links it placed. Then it starts the next page. To change course, just say so.

How many pages per session

Plan on 2 to 3 pages per Claude Code session. Past that, quality drops before the session runs out. For a longer batch, open a new session in the same folder and type the same line: it continues with the next page and never rewrites one already done.

How an article is made

Five steps, in this order, for every page:

1. Gather the evidence. It collects what the page may rest on: your own pages that cover the subject, the pages ranking for it, the questions the page must answer, and a sample of your writing to match. It also writes the page's brief.
2. Write. The draft answers every question the page owns, takes a clear position, and places the internal links the plan lists.
3. Check. Every sourced claim must point to a real source. Every price, rate, date or duration must be sourced. Anything taken from another business must name that business. A draft that fails does not go further.
4. Remove the machine tone. A second pass rewrites for rhythm and variety, measured against your own pages, without touching any fact, figure or source.
5. Check again, then publish. The check runs a second time on the rewritten draft, and only a passing draft becomes the final file.

How it uses sources

What needs a source:

• a price, rate, date, duration or anything a reader spends money on;
• anything about your business: hours, services, areas, credentials;
• anything taken from another business, which is also named in the text.

What does not: practical advice, the steps to follow, trade-offs and how to decide, mistakes to avoid, and reasoning from facts already on the page. These carry no figure and name no business, and a page without them is unfinished.

What it covers. The headings of the pages ranking for the subject show what a complete answer covers. The article covers those decisions in its own order and shape. It never copies another page's structure and never takes a fact from one without naming it.

Who it may name. This follows your site:

• A shop, SaaS or local business never names its competitors, except on a comparison page the plan asked for.
• A blog that sells (bookings, products or a service) never names businesses selling the same, and may cite guides, press and official sources.
• A blog that does not sell may cite other publishers by name.

One source never carries a page. When a single business supplies most of an article's sourced claims, or was already the source of a previous article, the check flags it, and the writer says why in the notes.

What it says about you is only what you published. It never describes your business as offering something none of your pages supports.

How it counts

• One page is one brief. The brief is used when a page starts. Your plan holds a number of briefs per period; see Getting started for how they reset.
• Free: resuming a page, rewriting a page already started, the checks, and the humanize pass.
• No data used by the writing itself. The research was bought when the plan was built.
• Guest-post articles written for an outreach campaign use no brief.

When your briefs for the period are used up, it stops before starting a new page and tells you the date the next ones unlock. Your plan and every finished page stay as they are.

Publishing

The article is content/<slug>.final.md: Markdown, with headings, lists and links, ready to paste into your CMS or convert with any Markdown tool.

• Internal links only point to pages that exist on your site today, or to articles already finished in the same folder. A page the plan will build later is named in the text but not linked, so a reader never lands on a missing page.
• Keep <slug>.sources.json. It lists every sourced claim with the sentence, the source and its address. Once the article is published without markers, it is the record that the article was sourced.

Every file it writes

All of these land in content/, one set per page:

• <slug>.final.md: the article to publish.
• <slug>.sources.json: every sourced claim: the sentence, the source it rests on, who published it and where.
• <slug>.brief.md: the page's brief, written before the draft: the questions it answers, the evidence available, who may be named, the suggested structure, the internal links and the voice to match. Useful if a person writes or edits the page.
• <slug>.md: the working draft, with source markers like [S012] still in place.
• <slug>.notes.md: the writer's notes: the positions the article takes and what each rests on, anything it could not source, the links it placed, and how it avoided repeating your other articles.
• <slug>.humanize.md: what the tone pass changed and why.
• <slug>.bundle.json: the evidence the article was allowed to use, with who owns each item. The check and the audit trail read it.

Once per folder:

• VOICE.md: how your site writes, measured from your own pages: sentence and paragraph rhythm, and which dash you type. The writer and the tone pass match it.
• briefs.ledger.jsonl: one line per page started, with the brief it used. A resumed session reads it so no page uses a brief twice.

What it does not do

• No ranking or traffic promises. It writes pages worth ranking; where they rank is decided by search engines.
• No word count target. A page is as long as the reader's decision needs, with evidence where a claim can hurt and advice where it cannot.
• Every article is unique. It reads your other finished articles first and avoids their openings, devices, structure and retold evidence.

When something goes wrong

"This period's briefs are all released." Your pages continue on the date given, or straight away on a bigger plan. Nothing already written is affected.

It reports "few sources" on a page. Your site and the ranking pages carried little on that subject. The page leans on advice and says in its notes what it could not source. Adding a page of your own on the subject, or naming a source you trust, gives the next run more to work with.

The check blocked a draft. The writer fixes the draft and runs the check again; you do not need to do anything. A page is never published with a blocking problem unless you explicitly ask for it, and then it is reported as forced.

A link from the draft is missing in the final file. It pointed to a page that does not exist yet, so it was removed. Build that page, then ask for the link to be added.

It says a page is not in the plan. Name the page as the plan's wave board writes it, or ask for "the next 3 pages" and let it pick in order.

A session ended mid-page. Open a new session in the same folder and type the same line. It finds the page in progress and continues without using another brief.