Ever tried translating hundreds of content guidelines into GPT-friendly versions? No? Smart move. I have, and it’s brutal work. But I’m not here to complain, though I’m exceptionally good at it. I’m here because I keep running into this question, in conversations and in my own head: WHY?
So let’s talk about it.
Every large corporate has one: the style guide. The content specialist’s bible, packed with rules for product spelling, sector-specific terminology, practical brand voice explanations, brand promises, examples showing exactly how to apply everything.
We’ve been working with these for years.
Enter AI…
Surely we can fit more in
The first idea that emerges (after some experimenting, naturally) is to dump all 50 or 100 or sometimes 150 guidelines from the style guide straight into the system promptEen prompt is de instructie die je aan een AI-model geeft zoals bijvoorbeeld ChatGPT. Het is hoe je communiceert met het systeem: wat je vraagt, hoe je het vraagt en... Meer. And honestly? That sounds smart. Feels seriously efficient.
Except, research (from OpenAI, for instance) shows that GPT models are not remotely reliable at following dozens of rules simultaneously.
And the more you cram into your prompt (system or otherwise), the greater the risk that GPT ignores the rules or blends them into some weird, tasteless text smoothie.
The result: output that still doesn’t meet all the rules AND feels completely flat, stripped of any trace of humanity.
Hint hint
And that’s your hint. Because people (content specialists especially) don’t walk around with all 100 rules in their heads either. More importantly, they don’t apply those 100 rules to all content, all the time, in exactly the same way. They do it contextually. A mortgage rejection needs a completely different emphasis than a birthday email. Right? Right.
And while GPTs keep getting smarter, they’re still pretty terrible at interpretation. Hit or miss. Sometimes almost magical, sometimes like a 90s raver chewing their jaw off #winkwink
Okay, okay, but we’re not giving up. What if we give the system smart sets with the right guidelines per content type, then tell it manually (or via a smart UI) whether it’s an email or a web page?
Bad news everyone!
That’s still not enough. An email can be light or heavy, compliance-heavy or informal. Without context (urgency, audience, risk) the output stays generic. And while I don’t expect GPT to immediately generate the-best-email-ever, every step away from that ideal is frustrating. It’s the start of endless tinkering, adjusting, checks, balances, and other time-sucks that mean I might as well have written it myself.
Worth noting: prompt research confirms this too. The more specific and context-rich your prompt, the better the output.
You can’t escape prompting
[Insert frustrated sigh] But what then? Because what everyone wants is maximum automation (not me, necessarily). No content specialists manually crafting prompts and pasting them into GPT. Hell no. We want systems that understand content types, know which guidelines apply in specific situations, know how to apply them. Systems that produce drafts that make the human-in-the-loopHuman-in-the-loop is de term die wordt gebruikt om aan te geven dat een mens AI-output controleert en goedkeurt voordat het wordt gebruikt of gepubliceerd. De term is wijdverbreid, maar klopt... Meer happy without 50 revisions.
There are roughly 2 ways to build that nuance into your GenAI instruction:
- Upfront in the system by loading up the system prompt or working with different agents handling subtasks
- Manual prompting for specific tasks, with human hands and eyes
In short: prompting is inevitable. The question is where. And the answer should be: where it makes sense. Yeah, vague. More concretely: add a manageable set of generally-applicable guidelines to your system, then design good prompts that add nuance based on task, tone, and type, and use them contextually.
System prompt: use formal address
User prompt: close this text with a concrete next step OR summarize the message in the first paragraph
Looking critically at your system prompt means content specialists can still add nuance without the system prompt drowning in guidelines. Because the more you cram in, the higher the risk of contradictory instructions and poor execution.
Yes, you could still build something with if Email, then specific-guideline-set logic. But then you’re making things unnecessarily complex. You’re building a mini-LLM inside your LLM that probably still won’t deliver the contextuality good content needs.
But how?
Still with me? Good. You’re a bit of a nerd and that’s the new cool. Now that’s clear, what’s next? Four things:
- Translate your style guide’s guidelines and rules into content guidelines and guardrailsGuardrails zijn beperkingen die je instelt in je prompt om te voorkomen dat een AI ongewenste, onjuiste of riskante output genereert. Guardrails werken als veiligheidsregels in je prompt. Ze vertellen... Meer. Remove ambiguity, vagueness, room for interpretation. Define tightly or quantify where possible (not “short sentences” but “sentences of max 8 words”)
- Make them easily findable and searchable. Add metadata and tags so you’ve got a usable collection of content guidelines that anyone can use (so you can quickly pull up all B1-level guidelines or everything touching compliance)
- Decide which guidelines you want manually included in prompts versus which can live at a higher level: system prompt versus manual prompts
- Manage your guidelines centrally, so when something changes (new brand voice, new legislation affecting content, etc.), everyone always has the latest versions
You’re probably thinking: “Who’s this ‘everyone’ who needs content guidelines?” Well, GenAI’s actual product IS content. Whether it’s website content, newsletters, service emails, chatbots, individual customer emails, CEO speeches. Every project using GenAI to generate content should use the same guidelines. That’s why we documented them pre-GenAI in some file on the intranet.
A shared, up-to-date version of those guidelines means your brand language gets spoken consistently across all communications. Even better than pre-GenAI.
Plus, a repository like this lays the groundwork for proper automation (now a best practice). I personally dream (sad, I know) about tech teams using an API to automatically pull the correct guideline sets (on the fly or at intervals) so we always know we’ve got the right set for the task at hand.
But that’s still a dream. Right now I’m thrilled that there’s a first version with GPT-proof content guidelines that are easily findable, divided into sets, tagged across relevant dimensions. All in an Excel that’s consumed my life for the past while.
Standardization vs nuance
Quick note: not every organization or team has the same GenAI needs. Legal or compliance departments benefit from highly standardized output. Same formulation every time, minimal variation. A heavier system prompt can work there, as long as rules are simple and unchanging.
But when you need variation, audience targeting, or nuance (marketing, HR, customer service) that approach backfires. You maximize quality by applying a selection of rules per situation. Modularity and context-aware prompting aren’t luxury, they’re necessity. At least if you want to mitigate risk, maintain brand voice, and keep your GenAI use transparent and manageable.
So… why?
Yes, it’s been an insane amount of work, but relevant work. Because a system prompt isn’t storage for your entire style guide. It’s a skeleton. Important, but insufficient.
The real intelligence is in making choices: which rules are relevant NOW? If you don’t organize that centrally and apply it modularly, you’re not automating quality, effectiveness, or efficiency. You’re automating chaos.
I need to catch my breath and not hear the word “guideline” for a while. But we’re not done. Because besides guidelines, there are other repositories worth building. More on that another time.
Don’t stress….it’s just me!
I’ve spent over 25 years working in content strategy and digital transformation, which means I’ve seen enough technology hype cycles to be skeptical and enough genuine innovation to stay curious.
Want to talk shop? Do get in touch!
I have a newsletter and it's bearable. Subscribe to read my (Gen)AI articles!
