Skip to content

Talk first, then three parts ​

A person can describe their work by talking. They do not fill in a form. The AI sorts what they said into tasks, shows each task as three parts, and asks only for what is missing.

The method is the same in two places:

  • In the person's own AI chat (Claude or ChatGPT connected to DutyGraph over MCP). Built as instructions in the work-discovery protocol.
  • In the client app, on a phone or a laptop, with hold-to-talk or typing. Built as Talk about your work.

The two rounds ​

mermaid
flowchart TD
    Q["Round 1<br/>One open question:<br/>'Tell me about your work'"] --> R["The person talks<br/>as long as they want"]
    R --> S["The AI sorts what it heard<br/>(see the reasoning steps)"]
    S --> L["A short list of tasks<br/>in the person's own words"]
    L --> C["Round 2<br/>One task at a time,<br/>shown as three parts"]
    C --> G["The AI asks for one<br/>missing part"]
    G --> C
    C --> OK["The person confirms<br/>and sends the card"]

Round 1: the open question ​

The AI asks one open question, for example "Tell me about your work. What do you do in a normal week?" The person talks or types for as long as they want. The AI does not interrupt with questions about fields.

Round 2: three parts for each task ​

Each task is shown as three parts, from top to bottom:

PartThe questionWhat it fills on the Task Card
1Where did you get it?The trigger, what comes in and from whom
2What did you do?What the person does, the steps, and how they know it is done
3Where did you send it?Who gets the result, and how

Each part shows what the person already said. A part that nobody described shows "not said yet". The person answers one missing part at a time.

A card is ready to send when the three parts are told from start to end and the card names who gets the result. Files, photos and screenshots can be added later. A card never waits for them.

The reasoning steps of the AI ​

These are the steps the AI follows between what the person says and what the person sees. They are written in the protocol that DutyGraph's server returns, so every AI app gets the same steps.

mermaid
flowchart TD
    A["Step 1: Sort each thing the person said"] --> B{"Is it work<br/>they do today?"}
    B -- "Yes" --> T["Step 2: Make a task<br/>title in their own words"]
    B -- "Looks recurring,<br/>not confirmed" --> ASK["Ask: 'Do you do this today?'"]
    B -- "Strategy, an idea,<br/>someone else's work,<br/>old, or unrelated" --> NO["Not a Task Card.<br/>Unrelated is never sent"]
    T --> F["Step 3: Put each fact in its part<br/>and mark it"]
    F --> M["said: the person's words<br/>guess: 'Suggested, please check'<br/>missing: 'not said yet'"]
    M --> N["Step 4: Repeat the names heard<br/>people, teams, systems"]
    N --> H["Step 5: Ask for one missing part.<br/>Hand-offs first"]
    H --> U["Step 6: Ask once about AI use"]
    U --> P["Step 7: Show the preview.<br/>Send only after a clear yes"]
  1. Sort. The AI puts each thing the person said into one class: work they do today, work that looks recurring but is not confirmed, strategy, an idea for the future, someone else's work, old work, or unrelated. Only work they do today becomes a task. Someone else's work becomes a hand-off on the person's own card, not a card.
  2. Name the task in the person's words. A cleaner title can be offered. The person chooses.
  3. Place and mark each fact. A fact the person said is recorded with their words. A guess by the AI is marked "Suggested: please check" and becomes a fact only when the person confirms it. An empty field stays empty. The AI never invents a value to fill a field.
  4. Repeat names. The AI says back the people, teams and systems it heard, so a wrong name is caught early.
  5. Ask for one missing part, hand-offs first. Who the work comes from and who it goes to are asked first, because they join this card to the next person's card on the Work Map. Then how often, how long, and what a new person would not know.
  6. Ask once about AI use. "Do you use an AI tool for any part of this? Which one, and is it a company account or your own?" The AI says that nobody is in trouble. The answer is recorded with the tools of the task. See The workflow coach and the AI use check.
  7. Preview and send. The server decides when a card is ready. The AI shows the preview and sends only after a clear yes from the person.

What the server checks ​

The AI follows instructions. The server does not trust that it did. These checks are code:

CheckWhat the server does
Is a value a fact?Only an answer recorded from the person, or evidence with its excerpt, is a fact. A guess is stored apart and never sent as a fact
Is the card ready?The server computes it from the required fields. A card that is not ready is refused
Did the person approve?Sending needs the person's yes after the preview. An answer to a question is never an approval
Does the person stop early?Every recorded answer is saved. The interview continues from the same place next time
Can anything be turned away?No. A person can always send raw text to DutyGraph, and the advisor sorts it later

In an AI chat, over MCP ​

The person connects Claude or ChatGPT to DutyGraph and says "Find my tasks" or "Let me tell you about my work".

  1. The AI calls start_work_discovery. The server returns the protocol with the two ways to start: from past chats, or talk first.
  2. After the person talks, the AI saves each task with save_discovery_candidates and records the person's own words with record_task_interview.
  3. evaluate_task_card_completeness returns the next questions and the progress. In apps that can show forms, DutyGraph's own interview form shows them.
  4. submit_task_card_drafts sends the cards the server finds ready, after the person's yes.

In the client app ​

Talk about your work is in My work and first in the Add sheet. It is on for every company (an operator can turn it off with SHOW_CAPTURE_FLOW=0).

  1. Talk. Hold the button and talk, or type, in any order. The words show on the screen and can be fixed with a tap. Done talking sends them.
  2. Found tasks. The AI sorts the words into tasks. What is not a task (an idea, someone else's work, something the person used to do) stays under Heard, not a Task Card. The person checks the names.
  3. The card. Each task shows its three parts and asks only for what is missing. I don't know and Save as is are always there.
  4. Lock in. When all three parts have words, the card is locked in and goes to My work as a draft for the person to check.

Intake is never blocked. With no AI connected, each thing the person sent becomes one task. Words not sent yet stay on the phone. Voice needs a speech connection (or the platform's transcription key); without it the screens open in typing and say so.

The client app's DutyGraph Agent then helps on each card: make it faster, finish it one missing part at a time, or time one run on the device.

INFO

Status. Built: the work-discovery interview over MCP with past-chat evidence, the said and guess rule, the readiness check and the approval step in code; from protocol 2026-11-13.1, the talk-first start, the three-part view and the AI use question as instructions to the person's AI. These three are instructions only: the server does not store "AI used" as its own field yet, and it records the answer with the task's tools. From November 14, 2026, Talk about your work in the client app (Talk, Found tasks, The card, Lock in) and the client app's DutyGraph Agent. Planned: an AI used field on a Task Card step.

Source: gitbook/product/discovery/talk-first-capture.md in the DutyGraph repository.

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