Generic AI output is a prompt problem, not a model problem. The model is doing exactly what it was trained to do: produce the statistical center of its training data. Your job is to pull it away from that center toward your specific voice. One prompt addition — a brand voice block — fixes most of it.
Why the model defaults to generic
The model was trained on billions of documents: blog posts, press releases, LinkedIn updates, corporate white papers. The average of those documents is corporate, hedged, and vague. Phrases like “cutting-edge solutions” appear millions of times in training data. They are the path of least resistance.
Without specific constraints, the model produces the average. That is what generic AI output is: the statistical center of the internet. It is not a failure of the model. It is the absence of a signal telling the model to go somewhere else.
The brand voice block
A brand voice block is roughly 100 words that goes at the top of every prompt. It contains four things:
- Your voice in three adjectives
- Your audience in one sentence
- Three things you never say
- Three things you always do
Here is the Avakata brand voice block:
Voice: Direct, practitioner, dry.
Audience: Performance-obsessed growth leaders and technical marketers evaluating agentic AI. Smart, skeptical, time-poor.
Never say: revolutionary, game-changing, seamless.
Always do: Lead with the point. Use specific numbers and timeframes. Explain the mechanism — why it works, not just that it works.
That block takes 90 seconds to write. It changes the output immediately. The model now has a target distribution to aim at instead of defaulting to the average.
Examples beat instructions
Abstract instructions are weak signals. “Be direct and specific” means something different to every document in the training data. The model averages across all of them.
Examples are strong signals. Include three samples of your best writing in the prompt — actual sentences or paragraphs you have already published. The model pattern-matches to examples faster than it follows abstract instructions.
“Write like this” followed by three examples outperforms “be direct and specific” every time. The examples show the model the target distribution; the instructions only describe it.
Keep the examples short — two to four sentences each. Long examples introduce noise. You want the model to extract the pattern, not reproduce the content.
Negative constraints are underused
Most prompts tell the model what to do. Negative constraints tell it what not to do. They are often more effective because they cut off the model’s default paths rather than competing with them.
Examples of negative constraints that work:
- Do not use the word seamless.
- Do not open with a question.
- Do not end with a call to action that starts with the word Discover.
- Do not use the phrase “in today’s” anything.
- Do not hedge with “might” or “could” when you mean it.
Each constraint removes one of the model’s high-probability default moves. Stack five of them and the output space narrows considerably. The model has fewer generic paths available and is more likely to produce something that sounds like you.
How to test your brand voice block
The test is simple. Take something you have already written — a paragraph you are proud of. Ask the model to rewrite it using your brand voice block. Compare the output to the original.
If you cannot tell which is which, the brand voice block is working. The model has matched your distribution closely enough that the outputs are interchangeable.
If the output is still generic, the block needs more signal. Add more examples. Add more negative constraints. Be more specific about what your audience already knows — a model that thinks it is writing for a general audience will hedge; one that knows it is writing for practitioners who can read a diff will not.
Iterate until the rewrite passes the blind test.
We send our Avakata brand voice prompt template to Field Notes subscribers. Get it at avakata.agency/contact.html.
The compounding effect of a good brand voice block
Once your brand voice block is working, every prompt that uses it produces on-brand output. You stop editing for voice and start editing for accuracy. That is a different kind of editing — faster, lower-stakes, and easier to delegate.
Across a high-volume content operation — say, 50 pieces a month — the time saving is material. Voice editing is the slow part. Accuracy editing is mechanical. A working brand voice block moves you from the slow part to the fast part on every piece.
The block also compounds. Each time you refine it based on a failed test, every future prompt benefits. It is a one-time investment with a per-output return.
If you want to build a brand voice block for your content operation and test it against your existing output, book a discovery call. We will walk through the process in 30 minutes.