How to Make Your AI Understand Your Messaging

Finally your whole marketing team is using AI? Congrats! How is it going? Oh, not so well? Now Content marketing's on ChatGPT, Growth is on Claude, and Social Media is on Gemini... And they describe the product like they work at different companies? And so you've started rewriting the messaging?

Why your LLMs write like six different companies

Because a model does exactly what you ask of it with whatever it is fed, then fills whatever you leave blank with the most average answer it can find. So if you feed it three solid case studies and your real USP, it’ll write more like you, but if you just say "make a LinkedIn post about our fintech platform," it’ll likely hand back the blandest, most generic, most vanilla post you have ever read. If other companies in your category or industry talk like that, AI will assume that’s right, and so that's what you get, regardless of which LLM you are using.

So when one person moans "our AI sounds generic" and another moans "our AI sounds off-brand," they're describing the same problem in two different hats. Generic is what you get when the LLM model has no clue who you are. Off-brand is what you get when the person prompting knows a totally different you than the teammate two desks over.

The point is, you can't prompt your way out of this. It's an input problem called lack of shared context. Luckily, it’s really easy to fix.

Better prompts won't fix this (and neither will a prompt library)

First off, let’s quickly walk through the two fixes we’ve been seeing circulate everywhere. Both feel clever, and both… flop for the same reason.

Fix one: tighten the prompts! Write better instructions, set up ChatGPT's custom instructions per person, maybe circulate a "how we prompt" doc. Lovely. It holds for about a week. Then someone tweaks their wording, a new hire never gets the memo, and you're back to everyone freestyling. Custom instructions live inside one person's account, in one tool. They don't hop over to the next teammate or, worse, the next platform.

Fix two: build a prompt library! Genuinely useful, too. But a prompt library without shared context just helps everyone crank out off-brand copy faster. Prompts are the questions. If every question still pulls from a different idea of who you are, congrats again, you've just scaled the confusion.

See what both of those fixes skip? They obsess over how people ask their questions, and ignore what the model actually already knows about you and your company. So even if your question is razor-sharp, if it’s aimed at empty context, it’ll still get you an average answer.

What you actually need is gloriously boring: one shared source of truth that every person and every tool reads from, whether that's Claude today or whatever shiny thing your team adopts next quarter. Nail that, and the prompts get easier too, because half the context you kept re-typing is already loaded in.

Which begs the question: what does that "source of truth" actually look like? And can you build one without blowing up your whole messaging?

Training your LLM with your current messaging

If your messaging is solid, you don't need to touch a word of it. You just encode it. In other, simpler words, make sure it lives in one place everyone on your team and all LLMs have access to it.

What we see most often is that some of it lives in a pitch deck, some more in separate docs, sometimes just the head of whoever wrote the last board update. Guess what. Your models can't read any of that! Your job here is really, really simple: round up the story you've already got and drop it somewhere a model can actually reach.

In practice, that means taking your ICP, positioning, value props and proof points, and shaping them into clean files a model can read. The kind of files your tools load and reference every single time they write. And look, Envy built a quick (and free!) tool that will help you put everything together:

By the way, the models aren’t the best at reading PDFs, PPTs and other fancy-looking formats, so best to keep it as simple as possible, in MD files.

If your messaging genuinely is a shambles, ten hands on it and no longer matching the product, then it's a different job. But luckily for you, it's exactly what the Elizabeth II workshops are for!

So what do these "clean files a model can read" even look like? Glad you asked.

What a shared context file actually is

"Context file" sounds like something only your engineer is allowed to touch, but it really isn't. It’s basically just a tidy, well-organised doc that tells any AI tool who you are before you ask it to lift a finger.

It's the brief you'd hand a sharp new hire on day one: who you sell to, what you actually do, why you're different, how you do it, and the proof to back it up. A context file is that brief, written once, in a format a model reads cleanly. Most teams save it as a markdown file (an MD), because it stays neat and loads into just about anything.

The good stuff is what goes inside:

• Your ICP, in proper detail: firmographic, psychographic, technographic
• Target audience profiles by persona and buying stage
• Core messaging: value props, differentiators, and the proof points behind them
• Positioning and your real USP, not the fuzzy elevator version
• Important: Add explicit instructions of things NEVER TO INCLUDE. For example, em-dashes, British English, abbreviations etc

Load that one file into Claude, into ChatGPT, into HubSpot, into whatever your team runs, and every last one of them starts on the same page. Literally the same page. And because it's a single file, updating your story takes one simple edit.

A few things to bear in mind:

• It’s really important though you always work with this one file instead of uploading multiple docs that might contradict each other. Your model will end up getting confused and producing stuff that doesn't necessarily make sense.
• Ideally the MD files should be saved in a shared Google drive and then all the LLMs connected to it. That way, any changes in the messaging get pulled across all LLMs, so you don’t have to chase individual teams to update their agents’ knowledge files.

Hand your LLMs one brain to share

You came here with one simple gripe: your AI writes like six different companies. Turns out the fix is smaller than the moan, yoohoo! Give every person and every tool one shared brief, and the output finally sounds like you, first try, across the whole team.

Elizabeth I exists for exactly that, and yes, it's actually free. It's a self-serve tool: feed it the messaging you already like, and it turns that into clean context files, plus dead-simple instructions to load them into Claude and HubSpot. You fill it in yourself. No agency hovering, no retainer lurking at the end, although we’d really appreciate it if you let us know the tool is super useful 😉

If your messaging needs more than a tidy-up… well, that's what the Elizabeth II messaging workshops are for. Either way you walk out with the files, the finished messaging and a process to keep it current. All yours to keep.

Your models are only ever as smart as the brief behind them, so go on, give them a good one.