Brand Voice Guardrails: The System Prompt Your Agent Needs

Voice drift is the default outcome of any AI agent left with a vague prompt, and “sound confident and friendly” is a vague prompt no matter how many adjectives you stack onto it. The fix is not a better adjective. It is writing brand voice the way you would write a spec: a fixed set of things the agent must and must never do, checkable against any single draft it produces.

A vague voice instruction cannot catch a bad draft, only a rule can

“Write in our brand voice” tells a model nothing it can act on consistently. Voice, to a language model, is a probability distribution over word choices, and a one-line vibe description barely nudges that distribution before the model reverts to its own defaults three sentences in. What actually holds is a rule the model can check a sentence against: does this contain a banned word, does this match the sentence-length pattern I asked for, does this topic require a human before it ships.

The skeleton has five parts, and none of them are optional

Here is the shape I use, built after watching an agent’s drafts slide off-brand more than once before I tightened the prompt around it.

Identity. One sentence, no more: who the brand is and what it is not. “We are direct and specific, never inspirational-poster vague” does more work than three paragraphs of adjectives.

Audience. Who is actually reading this. A prompt written for “everyone” produces writing aimed at no one in particular, which reads as generic by default.

Banned phrases, with substitutions. This is the part people skip and the part that carries most of the weight. List the exact phrases the brand never uses, and for each one, what to write instead. “Delve into” becomes “look at.” “Seamless” gets deleted, not replaced, because the sentence usually does not need the adjective at all.

Platform norms. A short line per channel on register, not tone. LinkedIn gets full sentences and no slang. X gets shorter lines, more fragments, less throat-clearing before the point.

Escalation triggers. Topics the agent flags and stops on instead of drafting through: pricing claims, competitor comparisons, anything legal-adjacent, any customer complaint that names a real person. The agent’s job on these is to say “this needs a human,” not to guess.

Encode what the agent should never touch, not just what it should say

The escalation list matters as much as the banned-phrase list, and it is the part most brand-voice guidance skips entirely because it is about restraint, not output. I write mine as a short list of triggers with one instruction each: if a draft would reference a specific customer complaint, stop and flag it instead of writing a response. If a draft touches a competitor by name, stop and flag it. If a number in the draft is a pricing or performance claim the agent cannot verify against the current facts it was given, stop and flag it. An agent that drafts confidently past all three of those is not a capable agent, it is an unsupervised one.

Register changes per platform, but the rules underneath do not

The mistake I made early on was writing a single prompt and expecting it to sound native on every network. It does not, because LinkedIn and X reward different rhythms even when the brand behind them is identical. My fix was small: one shared identity and banned list, then two or three lines per platform on register. LinkedIn drafts get full sentences, a professional register, no slang. X drafts get compressed, more fragments, faster to the point. The banned words and the escalation triggers do not change. Only the shape of the sentence does.

Test the prompt against edge cases before a real account touches it

Before I trust a prompt on a client account, I run it against a handful of deliberately awkward inputs: a customer complaint, a pricing question, a mention of a named competitor. Then I read the drafts as if I were the client seeing this land on their feed for the first time, not as the person who wrote the prompt. If any draft would make me wince, the prompt needs another pass. This is a five-minute check that catches most of what an escalation trigger would otherwise have to catch live.

Iterate the prompt, do not rewrite it from zero every time

Once the skeleton is in place, most fixes are small edits, not rewrites. A draft slips through with a banned phrase, add it to the list with a substitution. A draft on a new platform reads wrong, add a two-line register note for that channel. Treat the prompt as a living document that grows a little sharper every time something slips, the same way you would patch a spec after a bug report rather than rewriting the whole thing from scratch.

The prompt above is my own practice, refined against real drafts going out under real brand names, and I would rather hand it to you as a starting skeleton than pretend a single template survives contact with your specific brand unedited. Adjust the identity line, adjust the banned list, keep the shape. I wrote about the broader discipline of catching automation’s rough edges before they ship in AI social media automation without slop, and the same rule applies here: the check has to happen before the post goes out, not after someone screenshots it.