Enterprise security and B2B technology clients
Method, guardrails, and the adoption problem nobody warns you about
Context: 2024–present
Scope: System design, voice guardrails, custom model configuration, rollout
The situation
Every writer on the team had their own prompt, so none of them produced the same kind of output.
There was no shared process and no prompt library. People had been told to experiment, so experiment they did, showcasing different results every time depending on who asked and how they asked it. All that variation meant the editing and review process was now a painful de-Frankensteining project. Adopting AI had managed to make the work slower.
Trust is really the only currency in brand marketing that matters, and credibility in security rests on sounding like one company with one point of view. A hundred people writing a hundred different ways undoes one asset at a time.
~1/3
faster from ideation through production to publication
A hundred people writing a hundred different ways undoes trust.
What I decided
Start from the voice and let the task sit on top of it. Most prompt libraries are task templates: write a blog, write a social post, write a launch email, write a nurture sequence. That approach produces passable, anonymous output. The system I built defines what the voice does and, more importantly, what it refuses to do, and the task-level instructions add to that.
Write the rules with their reasons attached. Each rule in the guide explains why the construction doesn’t work, which makes it teachable and lets a writer exercise judgment in the cases the rules do not cover.
What’s in it
Banned words and phrases. Around twenty terms the industry has worn past the point of meaning, most of them endemic to security and B2B marketing: landscape, seamless, robust, unlock, game-changing, best-in-class, and leverage used as a verb.
Banned constructions. No “it’s not X, it’s Y”; you can just say something is something. No opening sentence that describes a mood instead of naming something specific. No conclusion stated as settled without evidence.
Structure and rhythm. Sentence length varies. No fake staccato, which is the most recognizable AI move there is. No defaulting to three examples.
Openings. Lead with a fact, a named source, a specific event, or a person. A draft needs to open with something concrete.
Evidence. A statistic supports an argument rather than replacing one. Projections get framed as projections.
An editing checklist. Eight questions to run before any draft is finished, ending with: does this sound like a person made specific choices?
[Screenshot slot: style guide excerpt]
Nothing publishes without human review
Where the line sits
Nothing is categorically off-limits to AI at the drafting stage, but nothing publishes without human review. How deep that review goes depends on what the asset is doing. Thought leadership, threat intelligence, research reports, and case studies get more reviewers and more cycles. Product announcements and podcast repurposing pieces get a lighter pass.
Video and podcast production is the exception. I will not deliver an AI-generated episode or video as a finished piece. AI can make sense inside that process (in transcript work, outlining, and cutdowns), but anything carrying a human face and voice needs a human behind it.
The hard part
Most teams want AI to make them faster. The argument that doing it properly upfront returns that time later is not persuasive to people under deadline. It gets harder once people have spent a year experimenting freely, because now it feels like something is being taken away.
What I’d do differently
I built the process and then rolled it out. I should have told people it was coming while it was still in progress, and asked what was working and what wasn’t before delivering something finished. The system would have been better for the input, and adoption would have been less awkward and uneven.
Marianne Chrisos 