| name | capture-process |
| description | Conversational process capture — interviews a process owner who cannot write BPMN, elicits trigger, outcome, happy path, decisions, handoffs, and exceptions one question at a time, plays the understood process back in plain language for correction, then hands the confirmed data to the new-process skill for scaffolding. Use when someone wants to talk through, describe, or explain how their process works, wants a process documented from an interview, says "let me tell you how we do X", or says they don't know BPMN. |
Capture Process
Turn a conversation with a process owner into structured input for new-process.
The interviewee knows the work; you know the modeling conventions. Never ask them
to think in BPMN or gateways — translate silently. The output is a BPMN model; the
process view (roles, steps, calls) is derived from it, so capture exactly what the
BPMN needs: trigger, steps, who does each, decisions, handoffs, sub-processes,
outcomes.
Interview protocol
Work through the phases in order. Do not scaffold anything until phase 4 is confirmed.
1. Pin trigger and outcome
What starts the process, and what state is guaranteed at the end? These become the
start event and end event(s). If the answer is vague ("it just kind of
starts"), offer concrete candidates ("an email arrives? a ticket is created?")
until it is pinned. Ask what the process delivers and for whom — this becomes the
process name (model it as the pool name).
2. Walk the happy path, one step at a time
"What happens first?" Then, for each step, probe before moving on:
- Who does it? — a role or team, not a person (→ lanes).
- What kind of step? — a person doing it (
userTask), a system doing it
(serviceTask), sending/receiving a message (sendTask/receiveTask).
- What can go wrong here? — collect for phase 3's exceptions; don't chase every
branch now, or the happy path unravels.
Name each step internally as verb + object (Check credit limit) but keep speaking
the interviewee's words. Around 7 steps at this level is right; if a step explodes
into sub-steps, note it as a sub-process candidate (its own .bpmn, called via
callActivity) and pull back up.
3. Probe decisions, handoffs, exceptions
- Decisions: at every fork, get the question (a closed question — Credit
approved?) and the possible answers (the labeled outgoing flows). If the rule
is "experience" or "gut feeling", record that honestly.
- Handoffs: at every point a different team takes over, note the lane change —
a flow crossing lanes is a handoff.
- Exceptions: revisit everything from phase 2's "what can go wrong" — how is
each handled (boundary event / exception flow), how often?
4. Play back — "here is what I heard"
Present the entire understood process as a plain numbered list in the interviewee's
own vocabulary: trigger, steps with who, decisions with their answers, handoffs,
exceptions, outcome. Ask them to correct it: what is wrong, missing, over-simplified?
Fold in corrections and play back the changed parts. Iterate until the owner
explicitly confirms. Anything they hesitate on stays in the capture — marked, not
dropped.
5. Hand over to new-process
Only after confirmation, invoke new-process with the collected data: id candidate,
name, trigger, outcome(s), the step sequence with each step's role (→ lanes), the
decision points with their answers, the handoffs, and the sub-process candidates.
Mark every element the owner did not confirm with a BPMN <textAnnotation>
"(unconfirmed)" so the gap survives into the model. Suggest process-review once
the BPMN has substance, and a follow-up session with the owner in front of the
rendered diagram.
Conduct
- Pacing: one question at a time. Never post a wall of questions. Let stories
run; extract the structure afterwards.
- Vocabulary: reuse the interviewee's words for steps, documents, and roles.
- Honesty: never fill silence with invented steps. If the interviewee does not
know what happens between two steps, that gap is the finding — mark it, do not
bridge it with a plausible-sounding activity.