After the draft, the agent asks what it left open

Why it asks

A first draft cannot know your approval thresholds or what happens when the payment provider times out. So after Create with AI produces a linked set, Runic sends the whole set to a reviewer prompt with one job: find what the draft does not answer yet, and ask only questions whose answer would change the diagram — a new .if, a validation step, a .fail with recovery, a .note threshold, a .back retry, a permission gate. It is told never to ask something the documents already answer, and to ask in the language the documents are written in. The questions are written by Claude Sonnet 5; each answer then goes through the ordinary chat route.

Questions bound to documents

The model answers in a strict line format: process name, topic, question. Topics are validation, decision, failure, timing, permission and data. Runic parses each line and binds it to a real document — exact name, then case-insensitive, then a containment guess, falling back to the main flow — drops duplicates and anything shorter than a question, and caps the queue at ten. Five to ten questions are requested, ordered by how much the answer changes the flow.

Asked one at a time

The chat opens with “The draft covers the happy path. I have N questions that would change the diagram…”. An “Open questions” panel appears above the message box with a progress count (“2 / 7”), a Skip button and a ✕ to close the queue. Each question switches the editor to the process it concerns, so you are looking at what you are answering about, and posts as “<process> · <topic>” followed by the question.

Answer in your own words. Runic wraps your reply with the open question and asks the agent to update the open document — the one the question switched you to — so the answer is reflected where it belongs: a validation step, an .if decision, a .fail with recovery, a .note threshold or a .back retry, keeping everything else intact. The result is an ordinary proposal: a diff on that process, previewed in the diagram, applied when you click Apply. The question is marked ✓ and the next one is asked. When the queue is done: “That’s the whole queue — 5 of 7 answered.”

Limits

The interview starts only after Create with AI produces a linked set of two or more processes in a signed-in session; a single generated document does not trigger it, and there is no button to start one by hand. It reviews up to twelve documents of 6,000 characters each, 30,000 characters in total, twenty runs an hour per account. The queue lives in the current session: questions already posted stay in the chat history, but the panel does not survive a reload. Skip marks a question skipped; ✕ drops all remaining ones with a note. An answer is a proposal — nothing changes until you Apply it.

How to use it

1. 1
   Generate a system

   Use ✨ Create with AI with a brief that describes an operation with several stages. When a linked set comes back, the chat opens on its own.
2. 2
   Read the first question

   The editor jumps to the process the question is about. The “Open questions” panel shows where you are in the queue.
3. 3
   Answer or skip

   Type an answer and press Enter, or click Skip. Click ✕ on the panel to end the interview.
4. 4
   Apply the diff

   The answer comes back as a proposal on that process. Review it, Apply or Discard, and the next question is asked.
5. 5
   Go deeper later

   After the closing message, keep asking in the chat — “add a retry to the credit check” works the same way.

Commands & shortcuts

Tips

• Short, concrete answers work best: “above 5,000 the CFO approves; otherwise the manager” becomes an .if with two branches.
• Skip questions that do not apply rather than answering “n/a” — a skipped question changes nothing.
• Apply lights up only while the proposal’s process is open. If the next question switched the editor to another process, go back to the first one before clicking Apply.
• The topic label tells you what kind of edit to expect: failure adds a .fail with recovery, timing adds a wait or a threshold note.

Questions