Write effective prompts

Prompt templates decide whether AI Engine gives your team specific answers or generic ones. Writing them is configuration work, not development, and it is often the change that improves answers most. This article collects what works. For the mechanics, see Create prompt templates and Add data sources.

Start from the people who do the work#

A template written for "assess client risk" without your company's own criteria produces something bland. The content that makes a template useful comes from the person who makes that judgement today. Ask an experienced rep or sales manager:

• When you open this account, what do you check first?
• What tells you a deal is in trouble?
• What should a new colleague notice that they usually miss?
• Show me an account you worry about and one you do not. What is the difference?

Their answers are the core of the template.

A structure that works#

1. Role and purpose. What the AI is doing and for whom. "You are preparing a sales manager for a quarterly account review."
2. The data. Name the data it works from: {context} for the record, {data:name} blocks under clear headings, {playbook} for company context.
3. The task. One job, stated specifically.
4. Constraints. What not to do, in plain words.
5. Format. Sections, bullet counts, length per section.

Example#

{playbook}

You are preparing an account manager for a renewal call.

RECORD:
{context}

OPEN CASES FOR THIS ACCOUNT:
{data:open_cases}

Write:
1. Account health in two sentences.
2. Up to three risks, each with the evidence from the data above.
3. Three questions to ask in the call.

Rules: use only the data above. If a section has no data, say so in one sentence.
Do not offer discounts or promise delivery dates.

Rules that hold in practice#

• One template, one job. A template that asks for six things gives weaker answers and runs out of output space more often than three templates that ask for two things each.
• Write in the language of the answer. Instructions in English tend to produce English answers. If your team works in another language, write the template in that language. When a language other than English is selected in the Actions bar, AI Engine also adds an instruction to answer in that language.
• Give limits per section. "Up to five items" and "two sentences" are followed more reliably than a total word count.
• Say what to do with empty data. A data source can return nothing. Without an instruction, the AI may skip the section silently or fill it with guesses. One sentence such as "If a section has no data, say so and move on" prevents both.
• Put totals in the query, not in the AI. Asking the AI to add up rows usually works and sometimes does not. A data source with COUNT() or SUM() gives a number you can check in Preview.
• Name data blocks precisely. A heading such as "OPEN OPPORTUNITIES PAST THEIR CLOSE DATE" is read correctly. A heading that is only a field name can be misread.
• Write guardrails into customer-facing templates. For drafts and meeting preparation, state what must not be said, for example discounts, admissions of fault, delivery promises or penalties. Rules that apply everywhere belong in the playbook's Compliance Guardrails. A template receives them only when it contains {playbook}.
• Refer only to data the template receives. A template that asks about a field that is not ticked in Agent Objects gets an answer built from whatever else it can see. That reads like an invented answer but is a configuration gap.

Leave room for the answer#

A long or structured answer needs output space. If a template's answers stop mid-sentence, or it sometimes returns nothing, check Max Output Tokens in the Org Profile before rewriting the prompt. Templates use that value or 3,500 tokens, whichever is higher, so raise it above 3,500 to give them more room. Answers in languages other than English use more tokens. See Org Profile settings.

Test before you roll out#

• Run the template on a record with plenty of data and on one with very little. Users find the thin record first.
• For templates with data sources, compare the answer with what Preview showed.
• Run it as a user with ordinary permissions, not only as an admin. Data sources run with the clicking user's access.

After go-live#

During the first weeks, check in the Activity Monitor which features people use most, and ask reps which questions they type into the chat again and again. The Activity Monitor shows usage by feature and user, not the questions people type, so you have to ask. Turn repeated questions into templates. A template gives a consistent answer and saves users from having to phrase the request well.