How to Write a Chatbot Persona Brief Before You Start Building

Replace descriptions like friendly and professional with specific rules a person or a generation system can apply identically.

Deriving rules from real responses

Collect twenty responses your best support people have written and read them for patterns, which produces far better rules than starting from brand adjectives.

The patterns are usually concrete, such as answering before explaining, using short sentences, or never apologising more than once in a message.

Making rules testable

A rule is usable if someone can look at a response and say definitively whether it complies, which excludes anything phrased as a feeling or an aspiration.

Maximum two sentences per point, no exclamation marks and never use the word simply are all rules that pass this test.

Listing words to avoid

A short list of words and phrases the assistant should never use is one of the most immediately effective sections in any persona brief.

Corporate hedging, minimising language such as simply or just, and any term customers do not use themselves are the usual candidates.

Setting length limits

State a default maximum response length and a hard ceiling, since unconstrained responses drift long and length is a top customer complaint.

Being specific here matters more than for almost any other rule, because it is the one most likely to be ignored without a number attached.

Rules for what the assistant will not do

Stating plainly that the assistant never speculates, never claims to feel anything and never guarantees an outcome is as useful as any positive rule.

These prohibitions are also the ones most likely to be violated by a generation system left without explicit guidance.

Specify how the assistant sounds differently in routine, frustrated and genuinely difficult situations, since one tone will be wrong for at least one of them.

Routine interactions

Define the default register for the everyday questions that make up most volume, since this is where the persona is experienced most often.

Efficient and light usually works, and calibrating the default here rather than for exceptional cases is the right priority.

Frustrated customers

Specify that responses become shorter and plainer, that apology is brief and singular, and that the route to a person becomes more prominent.

Write an example, since this shift is easy to describe and difficult to execute without seeing what it looks like.

Genuinely difficult situations

Bereavement, hardship, health and safety matters require the assistant to become quieter, briefer and quicker to hand over, and this should be written explicitly.

Any business whose customers might contact them in these circumstances should include this section rather than assuming general guidance covers it.

For each rule, show a response that follows it alongside one that does not, using the same question.

Why pairs work better than rules alone

Two versions of the same response make the rule concrete in a way no description achieves, and reviewers immediately understand which they prefer.

Pairs are also directly usable as examples for a generation system, which means the effort serves two purposes.

Covering each register

Include at least one pair for routine, frustrated and sensitive situations, since these are the three registers the brief has defined.

Six to eight pairs in total is usually sufficient, and a brief with twenty becomes something nobody reads.

Using real questions

Base the examples on questions customers actually ask, taken from transcripts, rather than invented ones that fit the point being made.

Real questions expose awkwardness that invented examples conveniently avoid.

The brief should be reviewed by support, by whoever owns brand voice, and by compliance where regulated topics are involved.

Support review

The people who write these responses daily will identify rules that sound sensible and do not survive contact with real customers.

They also know which phrasings produce follow-up questions, which is information no brand review will surface.

Brand and compliance

Brand review keeps the assistant consistent with everything else the business says, and compliance review matters wherever regulated topics are discussed.

Reviewing the brief is considerably easier than reviewing a live system, which is the practical argument for writing it first.

Reading examples aloud

Reading the paired examples aloud during review exposes stiffness that looks acceptable on the page, which is a crude technique that works reliably.

It also settles disagreements about tone faster than discussion, since the difference is immediately audible.

Keep It Short and Current

A persona brief that runs to ten pages will not be used, and one that stops matching the deployed assistant becomes misleading.

Two or three pages

Identity, rules, register shifts and examples fit comfortably in a few pages, and anything longer is usually background nobody applying the brief needs.

Length is a reliable predictor of whether a brief gets read, which makes brevity a functional requirement rather than a stylistic preference.

Updating from real conversations

Transcripts show where the persona is not working, particularly responses that read badly in context or fail to shift register when they should.

Reviewing these quarterly and updating the brief keeps it connected to the deployed assistant rather than describing an early intention.

Naming an owner

Without a named owner the brief is written once and never revisited, which is how the deployed assistant and the documented persona diverge.

The owner needs no authority beyond keeping the document accurate, which is a modest ongoing commitment.

Chatboq applies tone rules, vocabulary, length limits and register shifts as configuration across every response rather than per flow.

One configuration from the brief

Rules from the brief are entered once and applied to all generated and suggested responses, so the persona is consistent without anyone reapplying it per flow.

That correspondence is what makes writing the brief worthwhile rather than an exercise producing a document and nothing else.

Register shifts triggered automatically

Frustration and sensitive topic detection trigger the register changes the brief defines, rather than depending on a designer anticipating every flow.

The paired examples from the brief serve directly as the guidance those shifts are based on.

Checking output against the brief

Reviewing a sample of responses against the documented rules identifies both configuration drift and rules that turned out to be impractical.

Rules consistently violated in practice are usually wrong rather than ignored, and the brief should change rather than the behaviour.

Starting from the support team's voice

Configuring the persona from patterns in your team's existing responses produces an assistant that sounds like the business rather than like an assistant.

It also shortens the review cycle considerably, since stakeholders recognise the voice as one already approved in practice.