Eight things we believe about content that gets cited.

I ran content for a handful of small companies before building this. Every month the same pattern: a spreadsheet of keywords, a freelancer or a model turning them into two thousand words each, a CMS filling up, and rankings that did not move. When the answer engines arrived the pattern got worse, because now the reader never even saw the page. They saw a summary with three brands in it and none of them was ours.

So the question changed. It stopped being "how do we rank" and became "how do we get named in the answer". That is a different problem with a different set of tools, and nothing on the market was built for it.

An answer engine does not need a thousand pages about your topic. It needs one page that answers the question cleanly, with sources it can verify and a structure it can lift a paragraph from. Publishing more does not help; it dilutes. Our plans have article limits on purpose, and the trial has five generations because five good ones are enough to see whether this works for you.

Look at what a language model does with a page: it looks for the sentence that answers the query and lifts it. Pages that open with three paragraphs of context lose to pages that open with the answer. That is why every article we produce starts with a TL;DR, uses questions as headings, and ends with an FAQ. Not because it is a trick, but because it matches how the reader on the other side, human or model, actually reads.

I tried the obvious things first. A single long prompt produced fluent text with invented statistics. Splitting it into stages helped expose factual problems but the prose read like a manual. Adding a voice file fixed the prose but the topics were still guesses. Only once research came from real conversations, the plan linked to real pages, and a fact-check stage could delete sentences, did the output become something I would put my own name on.

Reprieve is that pipeline, wrapped in an interface, with the parts that need a person left to a person.

When you add a site we read it and write two files: COMPANY.md, what you do and for whom, and TONE.md, how you say it. Both are plain markdown. Both are shown to you. Every stage of the pipeline reads them, so if an article sounds wrong you can fix the file once instead of fixing every draft.

Keeping the voice explicit also keeps us honest. There is no hidden brand model that we tuned; there is a document, and you can disagree with it.

A topic is something a marketer decides to cover. A signal is something a buyer said. The second is more useful, because it comes with the words they used, the objection they had, and the thread where the argument is still going on. Our research stage collects signals: forum threads, questions phrased the way people phrase them, and what the current top answers are missing. You choose which ones deserve an article.

The pipeline pauses three times on purpose. Before writing, it asks you three short questions: which angle, which reader, which stance. After the draft, you read it. Before publishing, you press the button. None of this is slow; the whole thing takes minutes. But an article that a person has read and chosen to publish is a different thing from an article a scheduler pushed out at midnight, and the answer engines can usually tell.

Outreach is a manual service that opens only when an operator is available. Replies must follow community rules and disclose affiliations; unsuitable requests are declined and their allowance restored.

Most content tools are dashboards with a text box hidden somewhere inside. We wanted the opposite: a text box first, with the dashboard nearby when you need it. Start with a thought, answer three questions, watch the stages complete, read, publish. Tracking is a table and a line. Settings are a page. Nothing animates that does not need to, and nothing needs a tutorial.

If you want to see the output before reading any more, there is an example article. If you want to see how each stage is instructed, the skill files are public.