1 | Define the user's task
State who wants to accomplish what, who receives the result and which workflow step follows. Do not use a broad ambition for an intelligent platform instead of a concrete task.
2 | Map existing software and information
List existing forms, tickets, knowledge libraries and collaboration tools, recording where data enters and leaves. New capability should assist where the work actually happens.
3 | Choose a useful output
The result may be a summary, material pack, suggestion awaiting confirmation or human hand-off. Not every task needs a chat interface, and not every result should automatically be written back.
4 | Design the exceptions
Discuss insufficient information, unclear goals, scope mismatches and system failures. Give each a clear route for clarification, scope changes, human handling or retrying.
5 | Build a compact evaluation set
Prepare common, difficult and boundary cases from real work, specifying an acceptable outcome for each. Evaluate against the task rather than only the fluency of generated text.
6 | Hand the method to the team
Deliver operating instructions, exception guidance, material-maintenance practices and an issue record. The workflow should remain understandable and improvable beyond one successful demonstration.
