8 Lessons for Writing AI Prompts as a Salesforce Admin

8 Lessons for Writing AI Prompts as a Salesforce Admin

By

Agents are turning into everyday assistants for your work, and AIforce gives those assistants a direct line into your Salesforce org. At Dreamforce 2026, my colleague Lay Durafe and I showed how Salesforce admins can use the same tools developers use to vibe code, and vibe admin instead.

The more you work with agents, the more the back-and-forth between you and the agent determines what you get out of it. That makes prompt writing a core skill for a lot of roles. If you’ve been to any of my talks on AI, you know I’m not a fan of the term “prompt engineering.” Engineering makes it sound like a job for a technical specialist. Prompt writing is writing, and the lessons you’ve learned from years of professional communication apply here.

So what goes into a good prompt? First off, you don’t need to write code. You write in plain language, the same way you write emails, requirements and user stories. For most admin requests, three to six sentences does the job. Name the exact objects, fields and Flows by their API names. Explain the business reason behind the change, say what you want back and set your guardrails, like working only in a sandbox. Then keep going, because each answer gives you something new to follow up on. A one-line prompt usually costs you several rounds of clarification, and a few extra sentences up front save you those rounds.

Let’s walk through eight lessons from professional writing and what each one looks like when your reader is an agent connected to your org.

Start with who and what

1. Know your audience

A memo for an exec reads differently from one for an engineer. With agents, you’ll get better results when you set the stage and your expectations early. Tell the agent who the project is for.

Something like “I’m the solo admin for a 200-person sales team, and our sales ops manager will maintain whatever we build” tells the agent how much to explain and how complex to get. A one-person shop and a team with a release manager need different answers to the same question.

2. Lead with the goal

Good business writing puts the ask up front. Prompts work the same way: “Write a validation rule that blocks closed opportunities without a close reason,” then the details.

Open with three paragraphs of backstory and the agent has to guess which part matters. Put the ask in the first sentence and let the details back it up.

Say exactly what you need

3. Be specific

“Make this better” fails in an email to a colleague and in a prompt. “Cut this to 150 words and drop the jargon” works in both.

For admin work, specific usually means using names. “Fix the opportunity rule” could mean any of the twelve validation rules on Opportunity. “Fix the Require_Close_Reason validation rule on Opportunity” gives the agent one place to look.

4. Give context and explain why

A colleague makes better calls when they know the reason behind a request. So does a model. “Keep it short because this goes in a mobile notification” gets you a better result than “keep it short.”

Admin requests carry a lot of hidden context. Maybe a field exists because finance pulls it into a quarterly report. Maybe a Flow runs at night because an integration loads data at 2am. If the agent doesn’t know that, it can’t account for it.

5. Define the output format

Professional writers ask what format the reader needs: a summary, a table, bullet points. Tell the agent the same thing directly.

Ask for “a table with field name, type and help text” and you get something you can paste straight into a requirements doc. Ask for “the metadata only, no explanation” when you plan to review the XML yourself.

6. Show examples

Style guides use sample sentences for a reason. One or two examples of the tone or structure you want beat a paragraph of description. Many agents let you upload files or add them to your project. When I write a blog post like this one, my first step is to gather examples and resources and have the AI read them.

For admins, an example can be as simple as an existing validation rule you like. Paste it in and say “follow this naming pattern and error message style.”

Keep it plain and set your guardrails

7. Use plain language

Clever phrasing and idioms confuse people skimming an email, and they make an agent guess too.

Skip the internal nicknames while you’re at it. Your team might call the renewal process “the Q4 crunch,” but the agent only knows what lives in your org metadata and what you tell it.

8. Set constraints

Editors hand writers word counts, deadlines and a list of things to leave out. When your reader is an agent, that list becomes your guardrails.

For an admin, guardrails look like this: work in my sandbox only, don’t touch active Flows, don’t change any profiles. Write them into the prompt so they’re part of the conversation from the start.

Put the eight lessons to work

Here’s what those eight lessons look like on a single request. A first pass at a prompt might read:

Make a validation rule for close reason.

That works, sort of. The agent will guess the object, the stage, the field name and the error message. Now try it with the lessons applied:

I’m the admin for a sales team of about 200 reps, and our sales ops manager will maintain this rule. Write a validation rule on Opportunity that blocks saving a Closed Lost opportunity when Close_Reason__c is blank. Leadership reviews loss reasons every quarter and a lot of them come in empty, so the error message needs to tell reps exactly what to fill in. Use the same naming pattern as our existing rule, Require_Next_Step_On_Commit. Keep the error message under 100 characters and in plain language. Return the formula and the error message only. Don’t deploy anything yet.

Yes, it’s longer. It takes an extra minute to write. And it gets you a rule you can review and ship on the first try, which beats three rounds of “no, I meant…”

Keep the conversation going

Most of the time, one prompt won’t hand you a finished solution. The first prompt gets you a starting point, and a series of follow-ups gets you to a solution or a plan for one. The conversation is where an agent earns its keep. Take this prompt:

Why is <Flow_Name> throwing an unhandled fault error on <Object>?

Once the agent returns a useful answer, you can follow up with:

Can you scan active Flows that might have the same error?

And if you want to keep the problem out of future builds:

Help me draft an email that describes this problem and how to prevent it in the future.

Treat the agent as an assistant from the first question to the final email, and you’ll get a lot more out of it.

Try it on your next request

The habits that make a clear email make a clear prompt, and you’ve been building those habits for years. This week, take one request you’d normally fire off in a single line and rewrite it with the eight lessons above. Then share your before and after in the comments.

Want a head start? In this post, we break down the Agentforce Vibes Admin Accelerator Kit that Lay built for our Dreamforce session, with 100 prompts you can copy and use.

Resources

100 Ready-to-Use AI Prompts for Salesforce Admins

100 Ready-to-Use AI Prompts for Salesforce Admins

In this blog post, we looked at how the writing skills you already use every day carry over to prompting an agent. Now let’s put them to work even more. At our Dreamforce 2026 session on Agentforce Vibes, Lay Durafe and I handed out something to practice on: 100 ready-made prompts for the jobs admins […]

READ MORE
introduction to AIforce in Action

Introduction to AIforce for Admins

In September, Salesforce announced AIforce, its strategy for bringing Salesforce to any AI model, agent, or interface where your teams already work. To see where it fits, picture the Salesforce platform as four layers. Data 360 brings your data together. Customer 360 holds your workflows, data models, and business logic. Agentforce is where agents get […]

READ MORE
Streamline Your Path to Production with Builder Central

Streamline Your Path to Production with Builder Central

Admins are managing more requests than ever, and let’s be honest, it’s getting harder to keep pace. Turning business ideas into production-ready apps and agents takes time. You’re constantly stitching together features, navigating complex data models, and toggling between Setup and different builders. At the end of the day, it means that you can’t keep […]

READ MORE