Note · 1 min read
How much freedom I give the LLM
At Zaly I constrained next-step selection with deterministic playbooks. The model retained freedom to write within instructions adapted to the lead.
Choose the action before writing
At Zaly, where I worked as a contractor, the LLM had too much freedom to choose the next step in a sales sequence. I redesigned that part with deterministic playbooks, selected from the lead’s situation and the organisation’s rules.
If the lead had visited the profile, for example, the system loaded the corresponding playbook. The model received influence instructions for that situation and wrote the message using the lead’s context. The organisation could edit the instructions.
Decide where creativity helps
The message could vary. Next-step selection stayed in filters and priorities that could be inspected and repeated. Changing the wording therefore did not also change the sequence logic.
My conclusion is that giving a model more freedom adds decisions that need evaluation. When writing an adapted message, variation can be useful. When deciding whether a lead meets a filter, I need a stable criterion.
Playbooks let me divide those responsibilities. I can study whether action selection works and, separately, whether the instructions produce good messages. That separation makes each part measurable and lets an organisation adjust its rules without relying on one prompt to solve everything at once.