Teach AI Your Brand Voice

A guideline says confident but approachable. An exemplar set shows it. Collect the pieces that worked and the pieces that missed, each with one line on why, plus the lines customers quoted back to you.

The rejected drafts matter as much as the published ones. A layer built only from winners teaches what to imitate; the losers teach where the line is.

Watch for the harvest that skips call recordings. Your customers' own words about their problem are the best voice material you own, and they are already recorded.

Could a person with no taste check this rule correctly, every time, with no argument? Yes means mechanical and it belongs in a linter. No means judgment and it belongs in the rubric and the exemplars.

Most voice documents fail because they mix the two and enforce neither. Be authentic as a checkbox does nothing, and a human catching punctuation by eye is spending judgment on something a script does perfectly, forever, for free.

The mechanical half becomes a check that runs before a person reads a word. Banned phrases spelled out, punctuation tells, casing, forbidden constructions, length budgets.

It reports the exact line and the exact rule, it exits non-zero on error, and it runs on outside drafts the moment they arrive, so that feedback is mechanical and immediate instead of a thread three days later.

Watch for the linter getting credit for quality. A clean lint means the tells are gone, nothing more, and a linter that starts being treated as a quality bar is a linter about to be ignored.

The judge scores what a linter cannot hold. Predictability, whether the claim is earned, register, voice match against the exemplars, and what the reader can do afterward. Blind, because a judge told the founder wrote it scores the founder.

Run it three times and take the majority. We measured this in our own system. The same draft, the same rubric, the same model, scored clean and slop in consecutive runs, four of nine clean on an unchanged piece.

A single run is a coin weighted by nothing you control.

Trust a reproduced finding over the verdict label. Two runs flagging the same sentence for different reasons means that sentence is the problem, whatever the headline said.

Score five pieces you know are good, five you know are bad, and two your team argued about, with the answers kept out of the judge's reach and the pass bars written down first.

If the judge cannot separate your known-good from your known-bad, the rubric is wrong, not the drafts. A judge that passes a known-bad is missing an axis. A judge that fails a known-good is enforcing something your voice does not believe, and left alone it will slowly flatten the writing.

A layer that lives in one person's cloud and gets downloaded, edited, and re-uploaded has as many versions as it has readers.

One source, one named owner, dated versions with a one-line reason, and retirements recorded loudly, so a replaced rule cannot keep being enforced from a stale copy.

Anyone outside your core team is the test case. If bringing a new writer to your voice takes a meeting and a follow-up thread, the layer is described rather than distributed. The brief should be a link.

What ships goes back in. A piece that landed becomes an exemplar, a piece that missed becomes a known-bad, and a correction you have now made three times becomes a rule, mechanical if it can be one.

A context layer that never grows is a document again, and documents are what sent the drafts back needing you.