KaziKazi

Do positive instructions work better than prohibitions in agent prompts?

For steering which tool an agent picks next, yes: naming the action and when it helps gives the model something to do, while a prohibition only rules something out. We rewrote many of the instructions in Kazi's production prompts this way. Hard authorization rules still stay as plain prohibitions.

Ken Wiltshire, founder of Kazi · Checked against the product on 2026-11-14

A real example

Our agent has a sleep tool and a way to launch background work. The old line was "Do not use the sleep tool. If a task can be done in parallel, do not sleep."

The replacement: "Overlapping independent work is how you finish more of the job in the same turn. Launch a subagent for work that can proceed on its own and keep going on the current task." Now the model has an action and a condition, not just a closed door.

Where prohibitions belong

"Never run destructive database commands without explicit confirmation" should stay exactly that blunt. It's a boundary, not a preference. Better still, enforce it in code so the prompt isn't the only thing standing in the way.

So: positive instructions for preferences and work patterns, plain prohibitions for authorization, and platform checks for anything that truly can't happen.

How to tell if a rewrite worked

Watch the tool calls, not the wording. A better-sounding instruction only counts if the agent makes different calls in the cases you care about. More examples in How positive instructions guide an agent's next tool call.