Your team probably has a brand voice prompt. It started as a paragraph pasted into a custom GPT or a Claude project. Six months later it runs two pages: a tone section, a banned-word list, “never use em dashes,” “don’t sound salesy,” “don’t mention competitors,” “avoid jargon.”
Every one of those rules went in because a draft did something wrong. And the drafts keep getting vaguer.
We call the result the Yoda problem. Yoda talks in short, wise-sounding lines and leaves you to work out what he meant. Riddle-like AI content does the same: every sentence sounds wise, none of them connect, and by the end of a paragraph neither you nor your readers can say what it meant.
The usual response is to add another rule. That makes it worse.
When AI content goes vague, the usual cause is your prompt, and every new rule makes it worse.
Short answer: AI content often sounds vague because your prompt blocks the plain way of saying things. When rules forbid naming causes, solutions, or the obvious terms, AI reaches for abstract phrasing instead. The fix is usually fewer, clearer rules: tell AI what it can say instead of only what it can’t, give it real context, and test what each rule costs.
How does a brand voice prompt get so long?
Each rule starts with a real complaint. A draft named a competitor, so you banned competitor names. Sales said a post sounded pushy, so you added “don’t sound salesy.” The CMO winced at “leverage,” so it went on the list.
Each rule makes sense alone. Stacked together, they leave AI no plain way to make a point. It can’t name the cause, can’t point to the fix, and can’t use the words your buyers use. So it writes around all three.
Why does that read like a riddle?
When AI can’t say the plain thing, it says something next to it. Here’s an example, built from the kind of brand prompt a finance software company might write. A prompt for a post on closing the books carries these rules: don’t make readers feel behind, avoid jargon like “month-end close” and “reconciliation,” don’t recommend a process, keep it inspirational.
The output:
Every close tells a story. When the numbers wait, clarity waits with them. The strongest finance teams don’t chase the calendar; they let it come to them.
That’s Yoda on month-end: three wise-sounding sentences and nothing to act on. Now loosen two rules: let it name the cause and use the term your readers use.
Month-end close runs late when invoices sit in inboxes until the last week, leaving finance to reconcile a month of activity in a few days. Code each invoice the day it arrives and reconcile the bank account weekly. By month-end, most of the close is already done.
Same topic, same tool. The second version stops talking like Yoda and tells the reader what to do, and each sentence builds on the one before.
Why readers notice it now
Vague writing isn’t new. AI made it cheap to produce at volume, and readers have learned what it looks like.
Wikipedia’s volunteer editors keep a public field guide, Signs of AI writing, built from thousands of AI-generated drafts. Near the top of the list: inflated claims about why a subject matters, and generic, marketing-style phrasing. Those are the habits a rule-heavy prompt produces. When AI can’t be specific, it fills the space with filler that sounds important. Readers have learned to spot Yoda on sight.
For a B2B team, that’s a real cost. A buyer who spots the pattern stops reading, and the post that took an hour to edit gets the same three seconds as one that took none.
What the companies building AI recommend
The companies building these models describe the same tension. Anthropic’s guide to context engineering advises writing instructions that are specific enough to guide AI, but not so rigid they try to script every case.
The same guide recommends giving AI the least information that still gets the result you want. A two-page list of things to avoid rarely meets that bar.
How can you tell your prompt has too many rules?
- Sentences sound like Yoda: profound, with nothing you could act on.
- Paragraphs read as separate ideas rather than an argument.
- Posts lead toward a conclusion and never state it.
- The copy avoids the words your buyers actually use.
- Drafts got vaguer after the last rule you added.
How do you fix vague AI writing?
- Put every rule in one place. Brand voice doc, banned-word list, the extra lines someone added to the custom GPT last quarter. Look for rules that overlap or contradict each other.
- Remove rules one at a time. Run the same request after each change, so you can see what each rule costs.
- Tell it what it can say. “Name the cause, but don’t name competitors” works better than “don’t assign blame.” We cover how to write rules AI can actually check in AI content guidelines your AI can actually follow, and what that looks like for top-of-funnel posts in What awareness stage content should and shouldn’t say.
- Check any step before the draft. If you generate an outline or brief first, its rules carry into the final copy.
- Give AI something real to say. Your brief, a buyer’s actual question, a customer quote. The more specifics it has, the less it fills space with filler.
This has a cost. Loosening rules brings back some of the problems they were written to stop, and you’ll likely re-add one or two bans after testing. Some rules that look fussy exist for a good reason, so test before you cut, and keep what earns its place.
What to give AI instead of more rules
Most brand prompts grow by subtraction: every bad draft adds a line about what not to do. The prompts that hold up grow the other way. They carry the positioning, the buyer, the proof, and the brief, so AI has specifics to work with instead of rules to work around.
That’s the idea behind the Foundation in VelocityEngine’s Campaign Operating System. Instead of a prompt that grows a line every time a draft goes wrong, your team loads its messaging, positioning, personas, proof, and voice once, and every campaign inherits it. We explain why a messaging doc on its own rarely holds in Is your messaging doc a suggestion or a system?
Voice works the same way. The Foundation sets voice as a handful of dials and limits instead of a ban list: how formal, how much humor, how strong a stance, and how often a device can appear (one metaphor per piece, say, rather than none). AI gets a range to work in, which is what keeps it from talking like Yoda.
Governance is what keeps it there. The rules live in one place the whole team works from, instead of in a dozen personal custom GPTs, so a fix made once carries into every asset and your message doesn’t drift as output grows. We still keep a short list of bans, and we test each one against real output before it stays.
Frequently asked questions
Why does AI content sound vague?
Often because the prompt blocks plain, specific language. When AI can’t name causes, fixes, or common terms, it substitutes abstract phrasing. Thin context makes it worse, because AI has few specifics to draw on.
Can too many rules make AI content worse?
Yes. Each rule may solve one problem, but stacked rules can leave AI no clear way to make a point. Test rules one at a time to see what each one costs.
What’s better than banning things in a prompt?
Specific permissions and real context. Tell AI what it can say, and give it the brief, buyer, and proof it needs to say it well.
How do I spot AI-sounding content?
Watch for inflated importance, stock phrases, and paragraphs that never land on a point. Wikipedia’s Signs of AI writing is a useful public reference.
The takeaway
When your AI content starts sounding like Yoda, the instinct is to add another rule. Try removing one first, and add context in its place.
Every new rule is a sign your message is already drifting. Get the free eBook The more campaigns you ship, the further your message drifts: where drift starts, and the early signal that shows it before your buyers or your board notice.
