Personas — brand voice for AI
Your customer doesn't want to talk to an LLM. They want to talk to your company.
A generic AI voice is indistinguishable from the next ten companies that asked the same model the same thing — and it quotes prices you never approved. A Persona is the voice profile of your AI: define it once, attach it to any agent, and every reply sounds like your most senior, most on-brand representative. Swap the LLM, swap the tools — the voice stays. Open source, Apache 2.0.
Live — a persona as an audience
Does this content fit its reader?
A persona isn't only a voice — it can also be a reader you measure content against. Pick an audience, paste some text, and get a fit verdict: a readability score plus a text-grounded AI review that flags the exact spans that miss, each with a rewrite suggestion. This runs live against the demo backend.
See the same content through a persona's eyes — a fit score, what lands, and the exact phrases that miss.
One model, three voices
Same question — “is the annual plan worth it?”
The same AI, the same model, the same question. Only the persona changes — and with it, how your company sounds.
The closer
EXECUTIVE · verbosity 2 · PERSUASIVE
"Worth it: you lock this year's price and get two months free. Most teams that switch recover the cost in 90 days. Want me to run the numbers for your volume?"
Sales chat on a corporate site
The advisor
TECHNICAL · verbosity 4 · INSTRUCTIONAL
A short paragraph walking through the month-by-month savings math, with each plan's limits laid out in bullets so the reader can verify it themselves.
Developer-relations or pre-sales
The friend
CASUAL · verbosity 2 · DIRECT
"Way cheaper yearly 😊 it's 16% off. Wanna go for it?"
In-app chat on a mobile product
How it works
Define once — govern everywhere
A persona is decoupled from the model and the tools, so it's the one place your brand voice lives.
- 1
Define the voice once
In Administration → Personas, describe who the persona is (a free-text system instruction), set tone, verbosity and language style, and pin the vocabulary you require and the vocabulary you forbid. An AI-authoring assistant can draft every field from a plain-language brief — nothing is saved until you approve it.
- 2
Attach it to any agent
Give an AI Agent a catalog of personas and star one as default. The voice is decoupled from the brain and the hands: switch the LLM, swap the tools — the persona keeps every agent it's attached to sounding like the same on-brand representative. A flow can even switch persona mid-conversation.
- 3
Govern in two layers
Forbidden terms are enforced twice — in the prompt, so the model never intends to say them, and again after the response, where any slip is masked as [***] before it reaches the user. It understands Portuguese, English and Spanish, and matches word variants and whole phrases — compliance auditors review one list, not every reply.
- 4
Measure the audience
Personas aren't only a voice. Turn one into a reader profile and score whether a document actually fits the people meant to read it — readability plus a text-grounded AI review. Persona Match runs that across many contents × many audiences on a schedule; Synthetic User Research interviews a cohort before you spend a real participant's time.
Far more than a system prompt
A Persona is something you operate directly — a voice, an audience, and a research participant, all reusable.
A voice for every agent
Attach a persona and the agent speaks in your tone, uses the words you want to be remembered by, and never the ones that get you in trouble. Talk to a persona directly on a shareable URL to approve the voice before it goes live — no agent required.
Audiences & content-fit
Model a target reader — reading level, domain expertise, vocabulary ceiling — and measure whether a page fits them. You get a readability score (no AI needed) plus a text-grounded AI review that flags the exact spans that miss, each with a rewrite suggestion.
Persona Match (N×N)
A reusable project that scores many contents against many audiences at once, filling a color-coded heatmap live. Re-runs on a daily or weekly schedule, re-reading only what changed — so your content and your audiences stay aligned. Exports to PDF.
Persona Dialogue
Put two or more brand voices in a room and let them debate a topic, turn by turn, live. It's the fastest way to hear where your voices diverge — a "diff of voices" — before you decide which to attach to which agent. Saved as reusable projects.
Synthetic user research
Interview a cohort of personas against a research script and summarize the findings by theme — with verbatim quotes and a "good enough" saturation signal that tells you honestly when more interviews stop adding anything. A discovery aid, never a substitute for real users.
Draft from a recording
Hand Turing five minutes of "here's how our ideal rep sounds" and it transcribes, classifies the style fields, and writes a system instruction — producing a persona draft for you to review and save. Derive, never auto-apply.
Why not just write a good prompt?
You can get part-way with prompt engineering. Reuse, enforcement and measurement are where a governed Persona pulls ahead.
| Turing Persona | Prompt engineering | Off-the-shelf chatbot | |
|---|---|---|---|
| One brand voice, reused across every agent | copy-paste | partial | |
| Mandatory & forbidden vocabulary, enforced | two layers | ||
| Post-response tone masking (PT · EN · ES) | |||
| Teach the voice by example (few-shot store) | manual | ||
| Live brand facts via MCP — no redeploy | |||
| Ground answers in your indexed content | manual | ||
| Big Five (OCEAN) personality control | |||
| Swap the LLM — the voice stays | rewrite | ||
| Audience personas + content-fit scoring | |||
| Persona Match — N×N content × audience | |||
| Synthetic user research at scale | |||
| Draft a persona from a voice recording | |||
| Configured by non-engineers, no code | partial |
“Prompt engineering” = hand-written system prompts maintained per agent; “off-the-shelf chatbot” = a hosted assistant with a single system prompt. High-level, as of 2026.
Teams ask us
The questions worth raising before you give your AI a brand voice — answered straight.
- What is an AI persona?
- A persona is the voice profile of your AI — a small, reusable bundle of decisions (system instruction, tone, verbosity, required and forbidden vocabulary, optional personality and knowledge grounding) that replaces an LLM's generic default voice with your company's. You configure it once in Viglet Turing ES and attach it to any AI Agent; the same persona keeps every agent sounding like the same on-brand representative.
- How is a persona different from a system prompt?
- A system prompt is text you hand-write into one agent. A persona is a governed, reusable object: it's enforced in two layers (forbidden terms are stripped both in the prompt and after the response), it teaches the voice by example from a few-shot store, it pulls live brand facts from an MCP server, and it's decoupled from the model — swap the LLM and the voice stays. It's also editable by non-engineers, and the same persona reuses across every agent instead of being copy-pasted.
- Can a persona keep the AI from saying the wrong thing?
- Yes — that's the point of forbidden terms. They're applied twice: in the prompt so the model never intends to say them, and again after the response, where any match is masked as [***] before the user sees it. It works in Portuguese, English and Spanish and recognizes word variants and whole phrases, so 'promotion', 'promotions' and 'promoting' are all caught. Compliance reviews one list instead of every reply.
- Do I have to re-write prompts when I change the LLM?
- No. A persona is the voice; the LLM is the brain and the tools are the hands. Because the three are decoupled, you can switch from one provider to another — OpenAI, Anthropic, Gemini, Azure OpenAI, local Ollama — and the persona keeps every agent it's attached to speaking the same way.
- How do I keep prices and promotions up to date without engineering?
- Point the persona at a Brand Context MCP server you control. On every conversation Turing calls it and injects the current facts as a 'brand facts' block, so marketing updates a promotion in their own panel and the very next message reflects it — no prompt edit, no redeploy. The same MCP feeds every persona at once.
- Can a persona tell me whether my content fits its audience?
- Yes. Model a target reader as an AUDIENCE persona (reading level, domain expertise, vocabulary ceiling) and score any document against it. You get a fixed readability score plus a text-grounded AI review that lists what fits, what doesn't — with the exact span, the reason, and a rewrite suggestion. Persona Match scales this to many contents × many audiences on a schedule.
- Is synthetic user research a replacement for talking to real users?
- No, and Turing is explicit about it. Synthetic research interviews a cohort of personas to sharpen your questions, surface likely themes and pressure-test an idea before you spend a real participant's time. A 'good enough' saturation signal tells you when more interviews stop adding anything, and every study reads as a discovery aid. Always validate with real users before you act.
- Do I need to write code to use personas?
- No. Personas are configured entirely in the admin UI, and an AI-authoring assistant can draft every field from a plain-language brief — a marketer can create one. A REST API (/api/persona and friends) is there for teams that integrate by code, but nothing about day-to-day persona operation requires it.
Give every agent one voice — yours
Self-host under Apache 2.0, define a persona in the admin UI, and attach it to any AI Agent. The complete, didactic walkthrough — with a worked example for every feature — lives in the Persona Book.