The reference document, the review pass, and the loop that close the gap between "AI-fast" and "sounds like us."

When a team tells me their AI content "doesn't sound like them anymore," the instinct is always to fix it in the prompt. Add more instructions. Paste in three example posts. Try a different model.

I've made the case before that the prompt window is the smallest step in the engine — most of what determines whether AI content lands lives upstream of the draft. Voice is the clearest example. You cannot prompt your way back to a voice the model was never shown.

The fix is a system the model carries into every draft. Three parts: a reference document, a senior review pass, and a loop. Here's how I build each one.

The style guide you wrote for human onboarding won't work here. It's written for people who already absorb tone by osmosis. The model has no osmosis. It needs the voice spelled out explicitly.

The version I build has six sections:

Context. Two sentences: who you are, who you're for. The model writes differently for a skeptical VP of Marketing than for a developer, and it can't infer the difference.

How you sound. Four or five adjectives — but each one paired with a "which means." "Confident, which means we state a position and don't hedge it." Decorative adjectives ("authentic," "bold") do nothing; operational ones steer output.

How you don't sound. The banned words, the anti-patterns, the registers you're never in. This is the section that does the most work, and the one almost everyone leaves out.

Three contrast pairs. An on-brand sentence next to a close-but-wrong one. One pair teaches the model more than a paragraph of description, because it shows the line instead of asserting it.

Structural defaults. Sentence length, how you open, how you close, how you format. The patterns a reader recognizes before they've read a word.

The proof bank. Recurring facts, real numbers, named (anonymized) examples the model is allowed to draw on — so it stops inventing tidy statistics you can't stand behind.

Once the document is in place, the draft step largely takes care of itself. The leverage moves to review.

Not line-editing. One question: does this sound like us?

This is where the seniority matters. A junior content marketer approves what's correct — grammatical, on-topic, readable. A senior strategist catches what's wrong in a way no checklist captures: the borrowed phrasing, the false confidence, the claim you can't back, the quiet regression to competent-but-generic. That judgment is exactly the layer AI can't supply on its own — the human who decides whether the output earns the company's name.

This is also why "what you'd never say" matters as much as what you would. Most off-brand drafts aren't wrong about the topic. They're wrong about restraint.

Here's the step that compounds, and the one almost no team runs.

When the review catches something off, don't just fix the sentence. Fix the reference document. Add the phrase to the "don't" list. Sharpen the adjective. Drop in a new contrast pair.

The draft is one piece of content. The document is the asset. Update the draft and you've fixed today. Update the document and the model starts every future draft from a sharper version of your voice. Within a quarter, the gap between "AI draft" and "sounds like us" closes — and keeps closing.

I keep the reference document inside Claude so it's in the workflow before the first draft runs — not pasted in by hand each time. Some teams encode it as a custom skill the model loads automatically. Either way, the principle is the same: the voice is built into the engine, not re-explained every session.

A document is only as good as the judgment that writes and maintains it. That's the part that doesn't come in a box. The reference doc encodes a senior strategist's taste; the review pass applies it; the loop preserves it. Buy a tool and you've bought one step. Build the system and you've built the thing that makes the step worth running.

If your AI content is fast but no longer sounds like you, that's almost never a prompt problem — it's a missing voice system. Reply "VOICE" and I'll send the one-page brand voice brief structure I use to set this up. If your content is shipping but not converting, that's the gap the three-week Audit untangles — reply "AUDIT" if that's where you are.

— Jill

P.S. Next issue: what to actually log in the loop — the two-minute post-publish note that makes the whole engine sharper over time. [Ships Tue June 23]