Run brand voice for more than one client and a specific kind of drift shows up. Client A's content sounds a little sharper this week than it did last month. Client B's assistant starts using a phrase you told it to avoid three sessions ago. Nothing dramatic happened to cause it. The instructions just live in enough different places that keeping them all lined up by hand, remembering which chat thread has the current version, stopped being realistic somewhere around client three.
This is the ordinary failure mode for a solo agency running brand voice across several clients' AI content, and it isn't a discipline problem. It's what happens when the thing holding your voice rules together is your own memory instead of a structure that holds them for you.
The fix underneath all of this isn't a better habit. Habits are exactly what break down first when you're context-switching between five clients in an afternoon. What holds is a place for the rule to live that doesn't depend on you remembering it correctly every single time.
Why brand voice drifts once you're running more than one client
Voice instructions tend to start as a message pinned at the top of a chat thread, or a paragraph pasted at the start of every prompt. That works fine for one client, because there's only one version of "sound like this" to keep track of.
Add a second client and you're maintaining two separate copies of that instruction. Add a fifth and you're maintaining five, each one slightly different from the others because you tightened a rule for client three last week and never went back to update client one. Nobody decided that client one should sound different from client three. It just happened, because there was no single place the rule lived.
The AI itself isn't the problem here. It follows whatever instructions are actually in front of it in that session, faithfully. The drift comes from the instructions getting stale, duplicated, or simply lost between one session and the next, not from the model forgetting what you told it.
The two-layer fix: one global voice rulebook, one override per client
The fix is to stop treating "sound like this client" as one instruction per client and split it into two layers that never need to be lined up by hand, because there's only ever one copy of each.
The first layer is a global voice rulebook: sentence length, banned jargon, how formal the copy gets, written once as a Blend at global scope, which makes it available across every Project. The second layer is a project-scoped override, one per client, holding only what's actually different about that client, a phrase to avoid, a slightly warmer or drier tone, an industry term they use instead of the generic one. Each client's working Blend references both layers. Tighten the global rulebook, and every client Blend that references it picks up the change the next time you Brew, without you opening five separate client files to make the same edit five times.
What belongs in the global layer vs. the client layer
The split isn't always obvious in the moment, so a rule of thumb helps: if a change should apply to every client, it belongs in the global block. If it's true for exactly one client, it belongs in that client's override.
"Keep sentences under 25 words" is global. "Client A never uses the word 'disrupt'" is an override. "No exclamation points in headlines" is global. "Client B's audience expects more technical vocabulary than everyone else" is an override. Get this split right once and it keeps holding as you add clients, because a new client only ever needs a new override written, not a rewrite of the shared rulebook underneath it.
Where ad-hoc fixes break down at scale
A pinned message or a saved doc feels like enough at two or three clients. It stops being enough for a specific, checkable reason: nothing connects that doc to the prompt you're actually typing into that day. You have to remember to open it, remember what you last changed, and manually retype the relevant part into whatever thread you're working in. Miss that step once and the client quietly gets last month's voice, in content that already went out before anyone noticed.
A referenced block doesn't depend on your memory the same way. It's part of the Blend itself, so it's there every time you generate, whether or not you thought to check a separate doc first. What still depends on you is the generation and copy step: PAL renders the text, and you're the one who copies it into ChatGPT, Claude, or whichever model you're running that client's work through. The consistency comes from the rules living in exactly one place. Making sure the finished content actually goes out the door correctly is still your call, every time.
For the full build, from the first global block to the finished per-client generator, A Reusable Brand-Voice Prompt System for Client Content walks through the complete system. The client brand-voice system in the gallery is the same structure already assembled: open it and see the global blocks and the per-client override directly, not just a description of them.
None of this requires moving every client onto one identical assistant. The whole point of the two-layer split is that client A and client B can still sound distinct, on purpose, while sharing the rules that were never supposed to vary between them in the first place.