How AI Orbit researches and writes a long-form article

By Thomas Lovas Lokoy · Published 16 October 2026 · 8 min read

This is a description of what happens between pressing "Generate" and seeing a finished article in your library. It is written from the code that runs the product, not from a feature list, so nothing below describes something planned or something that only works on a demo account. Where a step depends on your plan, the plan is named.

What you hand over

The form on the "New article" page asks for a small number of things:

• A topic. Between 3 and 300 characters. A sentence is enough; the research step does the expansion.
• A voice. Five personas are available to every account, including The Futurist, The Satirist, The Sentinel and The Maverick. You can also switch the voice off, in which case the writer is told to use a plain, neutral editorial register with no persona styling.
• A length. A slider from 300 words up to your plan's ceiling, in steps of 100. It starts at 1,000 words, or at the ceiling if the ceiling is lower (Free accounts top out at 800).
• A funnel position (Starter and above). Awareness, Consideration or Decision, with Consideration selected by default, and an optional "what should the reader do next" line of up to 200 characters.
• Keyword research (Pro and above). A switch that is on by default once your plan includes it.
• A model. Sonnet costs 1 credit. Opus costs 3 and is available from Pro.

That is the whole brief. There is no prompt to write.

Step one: the charge comes first

Before any model is called, the credit ledger debits the article: 1 credit for Sonnet, 3 for Opus, or the day's free article on the Free tier. The debit is a single row-locked database operation, so two simultaneous requests cannot both spend the same credit.

Everything after that point is wrapped in a refund. If the research call fails, the writer returns nothing usable, the connection to the model drops, or the save fails, the exact amount that was charged is returned, and the article is marked failed with the reason. The message you see says so plainly: "Your credit was not used." If a generation is cut off without an error, for example because the browser tab was closed mid-run, a separate check closes any article that has been running for more than 20 minutes and refunds it the next time the library is opened. Nothing stays "generating" forever, and nobody pays for an article that was never written. The companion tutorial on credits, plans and limits covers the ledger in more detail.

Step two: reading the live web (Pro and above)

With research switched on, a separate call goes to Claude with the web search tool enabled and a hard cap of six searches. It is instructed to search the topic, search the question forms ("how", "why", "is it worth it", "vs", "best for"), and note what the top results do not cover. It returns a keyword brief with five parts:

• one primary phrase,
• up to eight supporting phrases,
• up to six questions searchers actually ask, phrased as they ask them,
• up to four gaps in the current top results,
• one sentence on the angle that would make a new article worth publishing.

The research instruction contains a rule worth quoting because it shapes everything downstream: the model may only report phrases it saw in a real title, heading, "people also ask" block or body text, and it must never state a search volume, difficulty score or cost per click, because it cannot measure them. That is why the brief panel on every article carries the line "No volumes or difficulty scores — we don't measure them, so we don't quote them."

You can run this step on its own before spending a credit: the "Preview the keyword brief" button on the form returns the brief without writing anything. Previews do count against the hourly cap shared with series and batch starts. If the research call fails for any reason, the article still goes ahead without a brief; research is an enhancement, never a reason to fail the piece.

Step three: the brief the writer receives

The writing call gets a long system prompt. The parts that matter to you as the person who will publish the result:

• Length is a commission, not a suggestion. The model is told that anything under 90 per cent of the target has not met the brief, and to plan roughly one section of about 350 words per heading. If it runs short it is told to add substance (a worked example, a real number, a counter-argument, a failure case), never to restate a point.
• Your voice. The chosen persona's instructions, or the neutral register if you switched the voice off.
• Funnel structure (Starter and above). Each stage has its own four-part shape. An Awareness piece opens on the symptom and closes with a next read, never a purchase; a Decision piece opens by confirming the decision and closes with the first action to take today. The reader must never see the words "awareness", "consideration" or "decision" anywhere in the text.
• Internal links (Starter and above). The writer is handed a list of your twelve most recently published articles, those with a public page, and told these are the only internal URLs that exist. It should work two to four of them into the body where a reader would genuinely want them, skip any that are not relevant, and never invent a path. If you have nothing published yet it is told not to write internal links at all, because a link to a page that does not exist is a broken promise.
• The rules. A banned-phrase list ("delve", "tapestry", "navigate the landscape", "game-changer", "it's not just X, it's Y" and the rest), exactly one H1 with ## sections, British or international spelling for a worldwide readership, and this sentence: "NEVER invent a statistic, a study, a quote, a person, or a source."
• Sources. The article must end with a Sources section listing only pages the model actually read, as links, and omit the section entirely if it read none.

When the keyword brief exists, it is appended with an instruction to use the primary phrase in the H1 and the first hundred words, to answer the questions as their own H2s where they fit, and to rewrite anything that reads like SEO.

Step four: writing, streamed

The article is streamed back. While it runs, the form polls the article row every four seconds and shows the phase: "Reading the live web…", then "Writing — 640 of 1,200 words" with a progress bar, then "Finding a cover image…". The word count on the row is refreshed at most every three seconds. The estimate shown beside the button comes from the length and the research switch: under a minute for a short article without research, up to four to seven minutes for a researched piece of 1,500 words or more. Keep the tab open; the request is allowed to run for up to five minutes.

When research is on, the writing call itself also has web search available, and every page it reads is recorded. Those pages appear under "Pages read while researching" on the finished article, alongside the Sources section the model wrote.

Step five: metadata, cover and save

The model's final line is a hidden comment containing a meta title (under 60 characters), a meta description (under 155), a one-sentence excerpt (under 200), three to six tags and a short visual description for the cover. The comment is stripped before anything is shown. If the model forgets it, every field is filled from the article itself, so a missing comment never breaks a page.

On Starter and above, that visual description is used to find a licensed cover image. Openverse is searched first, restricted to licences that permit commercial use and modification; Wikimedia Commons is the fallback, with anything marked fair use or non-free skipped. A candidate is only accepted if its title shares at least two meaningful words with the query, on the principle written into the code: a wrong picture is worse than none. The chosen image is shown with its title, creator, licence and a link back to the source page.

Finally the row is saved: the H1 becomes the title, the word count and a reading time at 225 words per minute are recorded, and the tags, keyword brief and sources are stored with it. The status flips to completed and you are taken to the article page.

What you should review before publishing

The product gives you the evidence; it does not sign off on it. Before you publish, check:

1. The word count against what you asked for. It is shown at the top of the article as "1,180 words of 1,200 asked for".
2. Every link in the Sources section. Open each one. The model is instructed never to cite a page it did not read, but the instruction is yours to verify.
3. Every figure in the body. Find it in the sources or remove it.
4. The keyword brief. If the angle is wrong for your audience, the article is built on the wrong foundation; regenerate with a sharper topic or a niche note rather than editing around it.
5. The cover licence. The attribution line links to the licence; make sure it suits where you are publishing.

Then publish to a public page, download the Markdown, or both. The tutorial on series, kinds and the public share link covers what happens next in the library.

What it does not do

It does not quote search volumes, because it cannot measure them. It does not promise a ranking. It has no in-app text editor: download the Markdown and edit in your own tools, or regenerate. And it does not spend a credit on a failure. See the plans for which steps your tier includes, or read how it works for the short version.