Transcript
Intro
We once gave a writing model a block of facts, headed: verified facts, use nothing else. It obeyed perfectly. In fifty five paragraphs of one batch, it wrote sentences like this one: the verified information does not confirm a dub. Very honest, and completely unpublishable. The model was fine. Our brief was the bug.
When a draft comes back wrong, the instinct is to fix the draft. Change a sentence, ship it, move on. It's the most expensive habit in AI content work, because the brief that caused the problem is still sitting there, ready to write the next hundred drafts the same way.
So today the brief is the product: we build one for our running example, take apart the ways briefs leak into the text, and put two popular claims on trial.
Why a one-line prompt fails
Start with what a model actually is. Google says it plainly: generative models don't retrieve facts, they predict a likely sequence of words. So a one-sentence prompt gets you the most likely page. Not the true one, and not yours.
Anthropic's own prompting guide says to think of Claude as a brilliant but new employee who lacks context on your norms and workflows. We'd put it more bluntly. A model is a fast junior with no memory of your site and no access to the truth. Anything it needs, you hand it.
And it isn't only the body text. Google's guidance says the same review applies to the title, the meta description, the structured data and the image alt text. Those are written from the brief too, so they come from the same pack, under the same rules. A perfect paragraph under an invented title is still an invented page.
Here's a one-line prompt for ReelRadar, our fictional film-trailer and where-to-watch site. Write a page about where to watch this film. Now count what that sentence leaves to luck. Which countries? Which services, checked when? Who is reading, and what do they need in line one? That's five coin flips before the first word.
What a complete brief carries
A complete brief has seven parts: the reader, the voice, the grounding facts, the output shape, the answer type, the floors, and the banned patterns. Let's build each one for that ReelRadar page.
Reader and voice first. Not everyone. Someone specific: a viewer in the UK deciding tonight whether this film is on a service they already pay for. The voice gets a plain description plus sample paragraphs you'd actually publish. Anthropic calls examples one of the most reliable ways to steer format and tone, and says to vary them, or the model copies one pattern.
Grounding facts are the heart of it. They come from your database or from sources you checked, and every one carries a date: which services carry the film in which country, subscription or rental, the runtime, the release year. The model writes from this pack and nothing else. If a fact isn't in the pack, the page doesn't need it.
Availability changes every month, so where the pack comes from matters as much as what's in it. A pack built from a stale table writes confident, wrong pages. So our pack builder refuses any row older than the freshness window we set, and leaves that country off the page instead of guessing.
Output shape says exactly what comes back: which sections, in what order, how many FAQ questions. And every FAQ question must be one the pack can answer. We check that before the brief goes out. You'll see why in a minute, and it's a good story.
Then the answer type, the part most briefs skip. Where to watch expects a list of services. When does it come out expects a date. How long is it expects a duration. Whatever the searcher expects goes in the first sentence, because that's the sentence a skimming reader actually sees.
Floors are minimums of substance, not of words. Google has said it has no preferred word count. So our floors count things a reader cares about: a checked service for every country listed, a date on every availability claim, and FAQ answers that each use a fact from the pack. Word floors exist too. They run last.
Last, banned patterns. We keep one list and it goes into every brief: dashes as sentence punctuation, stock phrases that announce a point instead of making it, inflated buzzwords, lists of three for rhythm, and one-line paragraphs for drama. Each one is harmless alone. Together they make every site sound like the same site. The full list is in the free download.
One habit makes bans work. Anthropic's guide says to tell Claude what to do instead of what not to do. So every ban in our briefs comes with its replacement. Not: no dramatic one-line paragraphs. Instead: paragraphs of two to four sentences, each carrying a fact.
Now get it out of the chat window, because a brief in a chat can't be versioned or tested. Ours is a text file in the repository. The grounding pack is JSON, built by a script straight from the database. Claude Code is the workbench: it runs the script, assembles the prompt, calls the model, and saves the draft right next to the pack that produced it.
Inside the prompt, every block sits in its own XML tag, which is also what Anthropic's guide recommends: the reader, the voice notes, the pack, the output shape, the bans. Instructions and material stay visibly apart. When a draft fails, you can see which block it misread.
Version the brief like code. Every change gets a one-line reason: which check failed, on which draft, and what the brief says now. Six months later, that log explains every odd-looking rule, and nobody deletes a rule they don't understand.
Before a brief runs a whole batch, give it a dress rehearsal. Pick five items that differ as much as possible: a famous film, an obscure one, one with no streaming at all, one with a messy title. Run all five and read every word. Most brief bugs show up in the weird items, never in the easy one you tested first.
How briefs leak into the text
Now the leaks. A model treats everything in the brief as material, including the parts you meant as instructions. Leak one is our opening story. The fact block said verified facts, use nothing else. And the suggested FAQ list included a question the facts couldn't answer: is there a dub?
So the model did exactly what we asked. It used nothing else, and it answered honestly by describing the gap. Anthropic's own advice on hallucinations is to let the model say it doesn't know, and for analysis that's great advice. In a page for readers, the confession lands in the prose.
The fix went into the brief, not into fifty five paragraphs. Never ask for a topic the facts don't cover. Don't give the fact block a label the prose can quote. Pair every do not claim X with do not mention X at all. And if the model must flag an unknown, give it a separate field, outside the text.
Here's what that separate field looks like. The model returns two things: the page text, and a short list called unknowns. The page text only says what the pack supports. The unknowns list says what it couldn't confirm, like dub availability, and that list goes to a person or into the next data pull. Never onto the page.
Leak two is instructions echoing as content. Openers like based on the information provided, or a sentence that politely restates your voice notes. One model we tested copied a line from our voice notes straight into a paragraph, like a student handing in the assignment sheet. Asking nicely doesn't reliably stop this. A gate does, and gates are the next episode.
There's a cheap way to catch it. Take the distinctive words from your own brief, headings and labels like verified, provided, pack and instructions, and search every draft for them. A word that only exists because of your brief has no business on a page for readers.
A database value is not prose
Leak three comes from your own database. In a generation pipeline we built, the brief took a title column as the canonical title. That column held the romanised Japanese name, because that's what the source catalogue stores. One audit found 204 drafts calling a show by a name its English readers don't use, and 165 did it in the very first sentence.
Here's the twist. The site's display code already cleaned that column, with a comment explaining the trap. The headline was right all along. The brief read the same column raw. Two layers read one value, and only one of them knew the trap was there.
That's the lesson that made our pipeline better: a database column is storage, not copy. Clean every value at every boundary where it reaches a reader, and count the brief as a reader. When you find one cleaning helper, search for every other place that reads the same column. That same week, a studio name stored with a lowercase legal suffix turned up in 130 drafts.
One warning before you start deleting things. That raw column usually exists for a reason. Ours lets the site's search find a show by its Japanese name. So we didn't purge it. We built a clean, reader-safe version where the brief is assembled, and left the stored value alone for the code that needs it.
Fix the brief, then regenerate
Now the receipt. Same ReelRadar page, two inputs. Left: the one-line prompt. Right: the full brief with its grounding pack. Both drafts run through the same five checks, live: answer type in the first sentence, every fact traceable to the pack, banned patterns, leak phrases, and floors.
Look at row two. The one-line draft can't pass the fact check even if every sentence happens to be true, because there's nothing to trace its facts back to. A draft you can't check is a guess with good grammar.
And when a full brief fails a check, nobody touches the draft. Read the failure, change the brief, regenerate. It's slower the first time. Over a batch it's much faster, because the defect is gone from every page after it, and the brief remembers why.
Verdicts
Two verdicts. Claim one: a better prompt fixes everything. Find the magic wording and the drafts come out right.
Wording changes style and structure. It can't add facts the model doesn't have, which is why Google still tells publishers to fact-check generated text before it goes live. In our pipelines, only grounding fixed the facts, and only a gate proved it. Verdict: placebo.
What a better prompt can do is real, just smaller. It can fix tone, structure and order. It can kill a repeated phrase. It cannot know which service carries a film in Germany this week. Only the pack knows that, and only a check confirms the draft used it.
Claim two: AI detectors decide your rankings, so every draft must be reworked until a detector calls it human.
Google's position since February 2023: it rewards quality however content is produced. In its words, using AI doesn't give content any special gains. It's just content. The spam policy goes after low-value pages made at scale, no matter how they're made. A detector score appears nowhere in either. Verdict: placebo.
Detectors are shaky instruments anyway. OpenAI pulled its own AI text classifier in July 2023 for low accuracy. At launch it caught 26 percent of AI-written text and wrongly flagged 9 percent of human text. A tool that misses about three in four, and accuses about one honest writer in eleven, should not run your editorial process.
Wrap up
The episode on one screen. The brief is the product: reader, voice, grounding facts, shape, answer type, floors and bans. Leaks get fixed in the brief. Database values get cleaned before they reach it. And no clever prompt or detector score rescues a page that has nothing to offer.
The brief template and our banned pattern list are free on the episode page at irongoo.com. Next episode: the gates that check these drafts, and how to test the gates themselves before you trust them.
