As a personal initiative, I’ve spent the last few months modeling a Salesforce-based tool that aggregates various CRM data points into multi-step flows, handing that structured information off to OpenAI via API for analysis and narrative generation. The criteria for each summary is driven by the end user, and the goal of the whole exercise is admittedly multifaceted. But at its core, I built it because I wanted a deeper understanding of what actually makes AI tick—and there’s no better way to learn than by designing something that forces you to interact with the model at a very granular level.

That said, the tool itself isn’t the focus of today’s discussion. I’ll unpack that architecture in detail another time.

Today, I want to focus on the true art form behind the scenes: the AI prompt—and how its defining properties shape the net result.

We’re going to talk about three characteristics that matter more than people realize: persona, temperature, and sentiment.
And this time, we’ll illustrate each with examples so you can see exactly how they warp—or elevate—the response.

---

Persona: The AI’s Costume Closet

Every prompt is a performance, and persona determines who steps onto the stage.

Assigning a persona changes the voice, priorities, and lens through which the model interprets your data.

Here’s the same input, interpreted through different personas:

Input Data:
“Opportunity ABC is at 40% probability, past due by 19 days, and missing executive alignment.”

CFO Persona Example

“Opportunity ABC shows elevated risk indicators: overdue by 1 days, lacking executive engagement, and stalled below 50% probability. Forecast accuracy requires reclassification or immediate senior intervention.”

Sales Coach Persona Example

“Opportunity ABC is within reach, but needs a push. Getting executive alignment this week could boost momentum and increase close probability. Let’s re-engage the champion and tighten next steps.”

Service Operations Persona Example

“The opportunity reflects process gaps—specifically stakeholder mapping and milestone discipline. Recommend a structured escalation and reinforcement of opportunity management best practices.”

Same facts. Completely different worldview.

Architect’s takeaway:
If persona isn’t explicit, the model chooses one for you. And much like a sandbox deployment done without a plan, surprises ensue.

---

Temperature: The AI’s Risk Appetite

Temperature controls how predictable or creative the model is.
Raise it, and you get originality. Lower it, and you get precision.

Let’s use the same input again:

Input: “Summarize pipeline risk factors.”

Low Temperature (0.1) Example

“Primary risk factors include stalled opportunities, low engagement from decision-makers, inaccurate close dates, and insufficient next steps.”

Straight facts. Zero creativity. Reads like compliance documentation.

Medium Temperature (0.5) Example

“Pipeline risk centers on stalled deals, weak stakeholder engagement, and inconsistent forecasting disciplines. Focusing on requalification and executive touchpoints will reduce exposure.”

More narrative, but still structured and business-ready.

High Temperature (0.9) Example

“Your pipeline is sending distress signals—stalled conversations, missing champions, and timelines that seem more optimistic than realistic. Shine some sunlight on neglected deals and re-energize executive sponsors to regain momentum.”

This version has personality. And metaphors. And opinions.
You asked for creativity—you got it.

Architect’s takeaway:
Temperature is a throttle.
Turn it up for ideation.
Turn it down for board decks.

---

Sentiment: The Emotional Styling Layer

Sentiment determines the tone—not the data, not the analysis—just the emotional packaging.

Using the same input:

Input: “The team missed its quarterly revenue target by 12%.”

Positive Sentiment Example

“While the team came in 12% under target, several emerging opportunities and improved late-quarter engagement signal strong momentum heading into next period.”

Optimistic, supportive, morale-preserving.

Neutral Sentiment Example

“The team finished the quarter 12% below the revenue target. Key contributing factors included delayed decisions, lower conversion rates, and limited executive engagement on strategic opportunities.”

Objective. No emotional cues.

Negative Sentiment Example

“The team fell short of the quarterly target by 12%, reflecting significant forecasting issues and inconsistent execution across critical deals.”

Sharply direct—occasionally useful, often dangerous.

Architect’s takeaway:
Sentiment is the difference between “We have an issue” and “Your career might have an issue.”
Choose carefully.

---

The Real Craft: Blending Persona, Temperature, and Sentiment

Together, these three properties create a signature style:

• Persona → What lens does the AI use?
• Temperature → How structured or creative is the response?
• Sentiment → How should it feel to the reader?

Consider how these combine:

Scenario: “Summarize customer churn risks.”

CFO Persona + Low Temperature + Neutral Sentiment

“Churn risk is elevated due to contract delays, reduced product usage, and lack of executive alignment. Recommend immediate account review.”

Customer Success Persona + Medium Temperature + Positive Sentiment

“A few customers show early signs of churn risk—primarily reduced usage and slower renewal conversations—but re-engagement is achievable through targeted outreach and refreshed success planning.”

Marketing Persona + High Temperature + Positive Sentiment

“Customer sentiment is shifting, but with the right nurturing and value storytelling, we can re-energize accounts and rebuild advocacy.”

This is why prompt design matters.
Not because AI is unpredictable—
but because you have far more control over its behavior than most people realize.

---

Final Thoughts

As AI becomes embedded in Salesforce workflows—automated summaries, guided insights, narrative reporting—the prompt becomes the new configuration layer.
Architects who master persona, temperature, and sentiment will consistently produce outputs that are aligned, actionable, and appropriate for their audience.

We’re entering an era where the strongest configuration tool isn’t a flow, a validation rule, or a permission set—
it’s the prompt.

Design prompts with with intention, and experiment boldly.