What belongs in a sales prompt pack

Most prompt packs are built around two things: who the buyer is, and how the agent should sound. Persona and ICP context, the buyer's role, their likely priorities, and the language they use for their own problem, gives an agent enough to write something relevant instead of generic. Tone and voice constraints, formal or casual, short or detailed, first person or company voice, keep the output consistent with how a real person on the team writes. Both are useful. Both get written first, because they are the most visible difference between a good draft and a generic one.

Two other parts matter as much or more, and get skipped more often: a real objection list, and explicit boundaries on what the agent must never claim. Both take more effort to write well. Someone has to think through what could go wrong, not only what a good outcome looks like.

An objection list is not a script of rebuttals. It is context that lets an agent recognize when a reply is an objection rather than disinterest, and respond to the specific concern instead of sending a generic follow-up. "We already have a system for this" is a different objection from "we don't have budget this quarter". An agent without a real objection list tends to answer both the same way. That reads as not having listened.

The boundaries list is the part that protects the business, not only the conversation. It should say explicitly what the agent must never claim: a compliance status the company does not hold, a guaranteed outcome or number nobody can back up, a price or discount the agent is not authorized to offer, a promise about a feature that does not exist yet. These are the mistakes that turn an awkward email into a real problem, a false claim a prospect can point back to later. They are far more consequential than a tone mismatch. This section deserves at least as much drafting effort as the persona and tone sections combined.

A well-written prompt pack makes the agent's typical output better and safer. It is still guidance to a model, not a hard constraint a system enforces. A model can misjudge an unusual case, misread an ambiguous reply, or drift from its instructions in a way the pack's author did not anticipate. Treating the prompt pack as the whole safety mechanism, with nothing checking what leaves the building, assumes a reliability no prompt can fully guarantee.

SalesCrew's agent policy keeps any external send, an email, a LinkedIn message, an SMS, in review by default, whatever prompt pack produced the draft. That is the deliberate second layer. The prompt pack shapes what the agent is likely to produce. The approval queue is the check on what gets sent. A boundary the prompt failed to enforce still has a chance to be caught before a prospect sees it. Nothing depends on the prompt getting everything right the first time, every time.