The Agency Playbook: Onboarding a New Client's Brand Voice in Under 10 Minutes

In short: You can onboard a new client's brand voice in under ten minutes. Point tone extraction at the posts they are proud of, add up to five PDFs or external URLs if their archive is thin, hand-edit the plain-text template for three minutes, then take sign-off from one proof post instead of a style-guide document that takes weeks.

Month one of a new retainer disappears into a document almost nobody reads twice. Most writing about ai brand voice wordpress setup explains what the tooling does, and almost none of it documents the procedure an account lead runs during the kickoff call. This is that procedure, on a clock.

Why does a brand style guide take four weeks when the writing already exists?

Branding engagements commonly run four to eight weeks end to end. The brand-guidelines phase alone gets quoted at two to six weeks, and the compressed "sprint" versions still take about two — the four-week case argued in startup branding timelines is the optimistic end of that range, not the average.

Most of those weeks are earned. Positioning research, competitor mapping, identity design: real work, real calendar time. But the branding process timeline bundles tone of voice into the same cycle as everything else, so the one deliverable your writers need on day one arrives alongside the logo files in week four. The content calendar does not wait that long. It starts publishing in week one, in whatever voice the drafting writer happened to have in their head that morning.

Intake makes it worse by asking the client to describe themselves. Look at any set of agency brand voice questionnaire questions and the same three show up: how would you describe your brand voice and tone, what phrases should we use consistently, what personality traits should come through. Clients answer "professional but approachable." Everyone answers "professional but approachable."

Good onboarding practice already says to keep the form under fifteen questions and ask only what the sales process failed to answer — the standing advice on client onboarding questionnaire best practice. Voice is the section to cut. The client has forty published posts written in their actual voice, and observed evidence beats self-description every time.

The month-one deliverable is not a style guide. It is a machine-readable voice profile that the publishing agents read on every draft.

What actually comes out of a 10-minute voice onboarding?

The contract is small on both ends. Inputs: the client's connected WordPress site, sampled by post type and taxonomy term, plus up to five PDFs and up to five external URLs when the site alone is not enough. What comes out:

• An editable plain-text template. Not locked JSON, not a black box — prose describing how this client writes, which you can rewrite by hand, revert to the generated version, or author from scratch.
• A Brand Visuals reference library, so image direction travels with the voice instead of living in someone's shared drive.
• A brand-style preview generator, which tests the look before anything reaches a live post.
• One isolated profile per connected site. That is the unit of separation across a roster.

Know what it does not do. It does not decide positioning, build a messaging hierarchy, settle pricing language, or replace legal review of claims. It captures how the client sounds, not what the client should be arguing. An account carries agents beyond the writing ones — a Chatbot AI Assistant Agent fields questions on the site, scheduling tooling runs the calendar — and voice is the setting the publishing side inherits.

Ten minutes buys the encoding step. It does not buy the relationship and it does not buy strategy. It buys the thing that unblocks the first draft.

What do you need from the client before the clock starts?

Nothing new. Everything on this list should already exist by the end of the sales process, which is the point — a kit, not another questionnaire:

• Admin or editor access to the WordPress site, provisioned before the call rather than during it.
• Three published pieces the client is proud of, sent as URLs.
• One piece they dislike, plus a single sentence on why. The most useful input in the kit, and the one agencies skip.
• Off-limits terms: competitor names, deprecated product names, the jargon their last agency overused.
• The name of the one person who signs off on voice.

With that in hand the ten minutes are structured rather than exploratory. Every block below names an action, an input, an owner, and the artifact it produces.

Run it during the call and narrate what you are doing while it happens. Clients who watch their own sentences get sampled stop asking for a style guide.

Minutes 0–3: sample the client's existing content by post type and term

The request lands in every kickoff, usually phrased exactly this way: train AI on my writing style. This is the block where it happens, and it happens by reading what the client already published instead of asking what they think they sound like.

1. Connect the site and open tone extraction. Choose the post type first. For most clients that is posts, but a consultancy's real voice often lives in case studies or a custom insights type.
2. Narrow by taxonomy term. If the three proud posts all sit in one category, sample that category — a term-level selection is the difference between a profile built on the founder's essays and a profile diluted by three years of release notes.
3. Exclude what the client has disowned. Legacy content from the previous agency, ghostwritten guest posts, anything published before the last repositioning. Ask directly: which categories should we not read?

Feeding a model approved examples and letting voice, tone, sentence structure and pacing come out of the text is the documented method here, and the practical write-up on training AI on approved writing samples lands in the same place. Abstract description produces abstract output. Twelve real posts produce a profile that knows this client opens with a number and never puts a colon in a headline.

Three minutes is generous for this. The selection is the work; the extraction is fast.

Minutes 3–6: the thin-content client (PDFs and external URLs)

Some clients arrive with nine posts, six of them written by a freelancer they fired. Sampling that site produces a profile of the freelancer. This is the branch most playbooks skip, and it is handled by the custom sample allowance: up to five PDFs and up to five external URLs, added alongside the site or in place of it.

Choose samples that share an author, not a topic. That is the whole rule. A founder's long LinkedIn essays and a whitepaper the founder actually wrote will produce a coherent profile even though one is casual and one is formal, because the habits underneath belong to the same person. Five on-topic PDFs written by five different people produce mush.

• Founder or executive essays, newsletters, long LinkedIn posts — usually the highest-signal source a young company has.
• A conference talk transcript or a podcast the client did. Spoken register sits closer to their real voice than their website copy does.
• A whitepaper or investor update they drafted themselves, for the formal end of the range.
• Sales emails that closed. Nobody thinks of these. They are the client at their most direct.

Skip the ghostwritten launch release, the agency-authored case study, and anything committee-edited. Thin or borrowed evidence is the first of the well-documented common brand-voice training failure modes: too little real material, and the output drifts toward generic competence. Three genuine samples beat ten borrowed ones.

Minutes 6–9: edit the template like an editor, not a prompt engineer

What comes back is prose. You can read it, disagree with it, and change it — no schema, no syntax, no delimiters to get wrong. Read it once at normal speed. Extraction is good at cadence and structure, and consistently weak on the things only you learned in the sales call.

1. Banned words. Pull the off-limits list straight from the pre-flight kit and write it in plainly: never use "solutions," never say "world-class," never name the competitor.
2. Product nouns, spelled the client's way. Capitalisation, spacing, whether the product takes "the" in front of it. This one edit prevents more revision rounds than anything else on the list.
3. Register by audience. If the blog talks to practitioners and the case studies talk to procurement, say so in a sentence rather than letting one profile average the two.
4. One explicit anti-instruction, taken from the piece the client disliked. "Never open with a question." "No numbered listicles." Negative constraints get specific in a way positive ones rarely manage.

Stop at three minutes. An over-edited template reads like a committee memo and produces exactly that, and if an edit overcooks it you can revert to the generated version and start again. Date the file and put a name against it. The template is a living document from here, and untracked edits are how a voice quietly stops matching the client six months in.

Minute 10: get sign-off with a proof post, not a PDF

The traditional approval cycle asks a client to review a twelve-page document describing writing, which is a request to imagine an outcome. Replace it with the outcome. Generate one real post on a real calendar topic in the new profile, and send it with a single question: does this sound like you, yes or no.

Binary on purpose. A yes closes voice onboarding inside the kickoff call. A no comes back with reasons, and reasons convert cleanly into template lines — "too formal" becomes a register instruction, "we would never say that" becomes a banned phrase, "too long-winded" becomes a sentence-length constraint. Two rounds is normal. Two rounds still lands inside day one, against a two-to-six-week alternative.

Give the review a 48-hour deadline and name the one approver from the kit. Voice sign-off by committee is how a four-week timeline reappears wearing a different hat.

Do the same for images before the first live post. The brand-style preview generator exists so you can test the look and settle it with the client in the same session, rather than discovering the mismatch on a published article. One proof post, one preview, and the account is ready to publish.

How do you run this across a whole client roster without voices bleeding?

Each connected site carries its own isolated profile, so cross-client contamination is not the interesting risk. The interesting risk sits inside a single client. One profile stretched across two audiences — practitioners on the blog, procurement in the case studies — drifts toward the average of both, which is nobody's voice. Split by site, or narrow the profile and give the second audience its own connected property.

Conventions that hold up past ten accounts:

• One profile per connected site. Never one per client, never one per campaign.
• A named editor per account. Not a team, a person. Edits carry their initials and a date.
• A re-run at month three. The client's own writing moves after a quarter of working with you, and the proof post takes ten minutes to redo.
• A logged reason for every hand-edit. Six months later, "why is 'platform' banned here" needs an answer.
• A quality pass before publish, without exception. Skipping review is the third documented failure mode, and it is the one that reaches the client's audience.

Voice is the first setting in an ai content agency workflow, not a deliverable parked beside it. Once it is signed off, calendar work is the next thing on the runway: the Editorial Calendar Agent takes the account from here, and every draft produced downstream inherits the template you edited in minute eight.

Key takeaways

• The month-one voice deliverable is a machine-readable profile, not a document. The document is the slow route to a worse artifact.
• Sample observed writing by post type and taxonomy term. Self-reported tone answers are the weakest input in the entire intake.
• For a thin archive, use up to five PDFs and five external URLs, chosen because they share an author rather than a topic.
• Budget three minutes for hand-edits: banned words, product nouns, register by audience, one anti-instruction.
• Sign off with one proof post and one yes/no question, then re-run the whole clock at month three.

FAQ

Can I really train AI on my writing style in 10 minutes, or is that just the setup?

Ten minutes is the encoding step, honestly labelled. It produces a usable profile and a proof post, which is enough to start publishing in the client's voice on day one. It does not produce positioning, a messaging hierarchy, or a claims policy — those stay separate work on a separate timeline. Expect to re-tune the profile at month three, when the client's own writing has moved and you have real published output to compare it against.

What if the new client has no published content worth sampling?

Use the custom sample path: up to five PDFs and up to five external URLs, added in place of the site archive. Pick samples that share an author rather than a topic — the founder's essays, a newsletter they write themselves, a whitepaper they actually drafted. A single-author set drawn from three sources produces a sharper profile than ten on-topic documents written by a committee.

Does this replace the client onboarding questionnaire?

It replaces the voice section of it, not the questionnaire. Keep the form under fifteen questions and drop "how would you describe your tone" — self-reported answers are the weakest evidence you will collect, and the client's published writing answers the same question in actual sentences. Spend the questions you save on what writing cannot tell you: the approval chain, claims that need legal review, competitors you must not name.

How do I keep 10+ client voices from bleeding into each other?

Each connected site carries its own isolated profile, so the cross-account risk is small. The failure that actually shows up is intra-client: one profile serving two audiences goes generic. Split by site, keep one named owner per profile, and re-run the proof post whenever the client's own voice moves. A leadership change or a repositioning is your cue.

Who on the team should own the voice template after onboarding?

One named editor per account, not the whole team. The template is editable plain text and can be reverted to the generated version, which makes it easy to change and just as easy to erode. Put initials and a date on every edit. When output drifts, you want a change log rather than a debate.

If you want the mechanics behind the clock — how tone extraction, custom samples, and the editable template actually work — the Brand Voice Agent page walks through each capability in detail.