Skip to content

The role AI persona ​

Each approved role profile can have an AI persona. The persona is a system prompt: the text an AI reads before it works for the person who holds the role. This page explains how DutyGraph builds that text and why it is built this way.

What goes in, and what comes out ​

mermaid
flowchart TD
    P["Approved role profile<br/>purpose, responsibilities, decision rights,<br/>recurring duties, measures, hand-offs,<br/>not decided yet"] --> C["Role compiler<br/>1 AI call when the profile is approved"]
    K["Kind of business<br/>and company name"] --> C
    C --> M["The mind of the role<br/>what it watches, what it believes,<br/>how it thinks"]
    C --> Q["First questions<br/>for the person in the role"]
    C --> S["3 task starters<br/>typical tasks, as ideas"]
    P --> A["Authority<br/>copied by code from the profile:<br/>do alone, ask first, work with"]
    M --> PR["The persona<br/>(system prompt)"]
    A --> PR
    PR --> AI["The person's own AI<br/>reads the persona, then reads<br/>the role profile and Task Cards<br/>from DutyGraph before it helps"]
    TC["Confirmed Task Cards<br/>steps, exceptions, done means"] --> AI

Two parts of the persona come from different places:

  • The mind is written by AI. It is how an expert in this seat weighs a case.
  • The authority is copied by code from the approved profile. AI does not write it and cannot change it.

The blocks of the prompt ​

DutyGraph puts the persona together in a fixed order. A block with no content is left out.

BlockWhat it saysSource
Opening line"You are the AI for the role role at company." Who the role reports to. "You work for the person who holds this role. You are not that person."Code, from the org chart
The roleThe purpose of the roleThe approved profile
What this role watchesOne question: the judgment this seat makes again and again, where a wrong answer costs the companyWritten by AI
What you believe3 to 5 beliefs of an expert in the seat. Each one says what it costs when it is ignored, and which role downstream gets the problemWritten by AI
How you think4 to 6 steps for any request. Step 1 is always "Find the responsibility this request belongs to. If it belongs to none, say so." The last step says when the work is doneWritten by AI. Without an AI connection, 5 fixed steps
What the role ownsThe names of the role's areas of responsibilityThe approved profile
Your factsFacts are in the role profile and the Task Cards, not in the AI's memory. Read them before you help. If there is no Task Card for a task, say soFixed text
Your authorityDo alone, Ask first or send to the role that decides, Work with, and NeverCode, from the profile's decision rights and hand-offs
Not decided yetThe open points of the profile, with "ask, do not guess"The approved profile
If you are askedSay that you are an AI when someone asks. The answer does not change with the rank of the person who asksFixed text

The Never list is the same for every role: never speak or sign as the person, never decide for the role, and never state a fact about the company that is not in the role profile or a Task Card.

Why it is built this way ​

The role, never the person ​

The persona models the seat. It names no holder, no personality and no history. Code also removes a holder's name when the profile shows one in brackets after a role title. The text stays correct when a different person takes the role, and it never speaks from one person's private answers.

Mind and ground are apart ​

The persona holds judgment. The steps, tools and numbers of the work stay in the role profile and the Task Cards.

The compiler tests each sentence it writes: if a duty or a task of the role is rewritten tomorrow, is the sentence still true? A sentence that fails the test is removed. A Task Card can then change without making the persona wrong, and the AI reads the current card each time instead of an old copy inside its prompt.

Authority comes from the approved profile ​

An AI that writes its own permissions can give itself more. DutyGraph does not let it. The Do alone and Ask first lists are the profile's decision rights, sorted by code. When a profile states no decision right, the persona says "Nothing is stated yet. Ask before you act."

Open points are named ​

An AI fills a gap with a plausible guess unless it is told where the gaps are. The persona lists what the company has not decided. The AI can then say "this is not decided, ask the role" instead of inventing an answer.

A changed profile cannot leave an old mind in place ​

DutyGraph stores a short mark of the profile that the mind was built from. When the profile changes, the mark no longer matches. The persona then uses the fixed thinking steps and the current authority until someone selects Build again. Current priorities are not part of the mark, because they change often.

How this helps the AI perform ​

Without the personaWith the persona
A general assistant gives a general answerThe AI starts from the judgment the seat makes most often
The AI does not know what the person may decideThe AI says "you decide this alone" or "the role decides this"
The AI invents a company fact that sounds rightThe AI must point to the role profile or a Task Card, or say the fact is not written
The AI treats every request as a new problemThe AI first finds which responsibility the request belongs to
The AI does not know when the work is good enoughThe last thinking step states "done" from the role's measures

DutyGraph has not measured these effects on customer work. They are the design intent of each block.

Where the persona is used ​

  • The person's own AI chat. When an employee connects Claude or ChatGPT to DutyGraph, the tools return their approved role profile and its persona. The assistant works inside it.
  • The advisor's AI. The get_role_persona tool and the dutygraph command line return one role's persona, role profile, confirmed Task Cards and the rules for answering as that role.
  • Copy and paste. The AI persona tab on the Role profile card has Copy. A person can paste the text into any AI chat.

What the persona cannot do ​

  • It does not approve, decide or commit anything. A person decides.
  • It does not open files that the person asking could not open in the app.
  • It does not replace the Task Cards. A role with no confirmed Task Cards can only be helped in general terms.
  • A task starter is typical for the kind of role. It is never saved as a person's Task Card. The person describes their own work.

INFO

Status. Built: the persona, first questions and task starters made in 1 AI call when a role profile is approved (needs an AI connection for the company); the plain persona made by code from the profile lines when there is no connection; the check for a changed profile; the AI persona tab with Build and Copy; the persona over MCP for the person's own AI and for the advisor's AI. Planned: a chat with the role inside the client app, the training modes for a new hire, and the workflow coach, which would use this persona.

Source: gitbook/product/architecture/role-ai-persona.md in the DutyGraph repository.

DutyGraph. Pages label each feature Built, In review or Planned.