How do I get good content from AI?

Close the specification gap before writing starts. Brief the model with the exact buyer question the page answers, hard pass-or-fail laws (no figures without a whitelisted source, structural floors, banned phrases), an anti-echo list of neighbouring pages not to restate, and then enforce all of it with an independent review.

What goes in an AI writing brief?

Seven parts that fit on one screen: the buyer question verbatim, the reader and what they already know, the shape a complete answer must take, laws stated as rejection conditions, an anti-echo list of what existing pages already cover, a closed source whitelist for any figure, and the pages the draft must link instead of repeat.

Why does AI content come out generic?

Because a model resolves every decision a brief leaves unstated by choosing the most statistically typical option — average audience, average structure, average position, average evidence. Generic prose is not a model failure so much as an underspecified brief being executed faithfully. Every real decision the brief makes removes one averaged choice.

Should an AI writing brief allow statistics?

Only from a closed whitelist of named sources, so any number in the draft either traces or fails — that turns review into mechanics instead of judgment. The strictest option, a flat no-figures rule, is underrated: it forces mechanistic reasoning over borrowed numbers, and those pages age better because nothing in them silently expires.

How to brief AI writers

Guide · SEO / AEO / GEO · 5 min read · last verified 2026-07-27

Reviewed before publication Editorial board Independent commercial review

In shortAn AI writing brief is a set of enforceable constraints, not a topic suggestion. The anatomy that works: one buyer question, laws not preferences, anti-echo lists, and closed source whitelists.

A brief for an AI writer is a set of enforceable constraints that turns a general-purpose model into the specific author of one page. It is not a topic suggestion, a keyword list, or a polite request for tone; it is closer to a contract, with laws the draft must satisfy and stated grounds on which it will be rejected. The distinction matters because of what models do with everything a brief leaves unsaid — and most briefs leave almost everything unsaid.

Generic output is the sound of averaged decisions

A language model completes an underspecified task by making the most statistically typical choice at every point where the brief is silent. Audience unstated: it writes for everyone, which reads as no one. Structure unstated: the standard skeleton of its training data, an introduction that promises, a tour of subtopics, a conclusion that recaps. Position unstated: the safest consensus view, held by nobody in particular. Evidence rules unstated: whatever plausible figures decorate similar text it has seen. None of this is malfunction. Asked to write an article about a topic, the model does exactly that — it produces the average article about that topic.

This reframes the craft. Generic output is not a quality problem to be edited away downstream; it is a specification problem to be closed upstream. Every real decision a brief makes is one the model no longer makes by averaging. The sections that follow are the decisions that earn their space — the brief anatomy Magrios applies to its own published library.

The spine: one buyer question

The single strongest constraint is naming the exact question the page answers — the question, not the topic. A topic licenses an averaged tour; a question has a discoverable answer, an implied reader, and a natural finish line, because the page is done when the question is answered. A named question also gives the eventual reviewer something concrete to verify the draft against, which a topic never does.

Two disciplines keep the spine honest. The question should come from evidence of what buyers actually ask — sales calls, support threads, what assistants get asked in your market, the same sourcing that goes into building a FAQ from real buyer questions — not from a brainstorm about what they might ask. And it should be one question per page, enforced: a draft that tries to answer three questions usually answers none well, and collides with the pages that own the other two. Selecting which questions get briefed in which order is its own planning layer, covered in planning a week of content from one research scan.

Laws, not preferences

Constraints work when they are phrased as pass-or-fail laws with consequences, not as stylistic wishes. A request to please avoid unsupported statistics is a preference the model will weigh against its instinct to decorate; a law stating that any number without a source from the whitelist below causes rejection is a condition it can actually satisfy. The difference in output is not subtle.

Working briefs carry a small set of these laws, stated flatly. Evidence laws: what may be claimed, and on what basis. Structural laws: a definition-first opening, minimum depth per section, no filler introductions, no conclusion that merely restates. Language laws: the banned-phrase list — the breathless opener about how fast everything is changing, the label that calls every feature transformative, and whatever additions your own batches have taught you to dread. Keep the whole set short enough to check. Ten laws that are enforced beat forty that are atmosphere.

The anti-echo list: what this page must not say

The least obvious section of a working brief is negative space: the neighbouring pages that already exist, and an explicit instruction not to restate them. Models given related briefs converge on the same explanations by default, and a library briefed without anti-echo instructions drifts toward a single averaged explanation republished under different titles. The fix is naming the neighbours — the definition of this concept lives at that URL; link it, do not re-derive it — so the new page builds on the library instead of repeating it.

This is also what makes internal linking mean something. A link to a page whose content you have just restated is decoration. A link that stands in for an explanation the draft deliberately omitted is structure, and it is the difference between a library and a pile.

The source whitelist

If a page is permitted to contain figures at all, the brief names exactly where they may come from — a specific named report, your own product documentation, your own published research — and states that everything outside the list is out, no matter how plausible. A closed whitelist is what turns the reviewer's evidence check from judgment into mechanics: any number either traces to the list or fails.

The strictest version is underrated: a brief that permits no figures whatsoever. It forces the draft to reason mechanistically about why things happen instead of dressing claims in borrowed numbers, and pages written this way tend to age well, because nothing in them silently expires when someone else's statistic does.

A skeleton that carries all of it

The parts above compress onto one screen:

QUESTION: the single buyer question, verbatim

READER: who is asking, at what stage, knowing what already

ANSWER SHAPE: what a complete answer must include to be done

LAWS: evidence rules, structural floors, banned phrases

ANTI-ECHO: neighbouring pages, and what NOT to restate from each

SOURCES: the closed whitelist, or a flat no-figures rule

LINKS: pages this draft must reference instead of repeat

Everything on the skeleton is checkable, which is the point: a brief is only as strong as the review that enforces it. The two form one loop — the gates a reviewer applies are the laws the brief declared, and the failures that keep recurring in review become the next revision of the brief. The review half of that loop is described in How to review AI-written content before publishing; together, brief and review are what carry research all the way to a page that survives a skeptical reader, closing the last mile from research to published page.