Summary  
This chapter explains how to structure an AI agent’s system prompt into four distinct parts—identity, tool guidance, output format, and constraints—to enforce consistent behavior.

General domain of usage  
Data analysis

You've used three pre-configured analyst agents already — one for the spreadsheet, one for the strategy doc, one for benchmarks. Same Claude model behind each. What made them act like different specialists was the system prompt — the persona.

Now you write your own.

## A Persona Has Four Parts

A good system prompt isn't a paragraph of vibes. It's a small set of named pieces, each doing a specific job:




"Always call spreadsheet_analyze before answering."

"Return: Headline, 3 Findings, One Recommendation."

"Never report a number without citing the row."

Write all four and the agent behaves consistently. Skip one and it drifts.

## What Drifts When You Skip a Part

* No identity — generic, hedging answers like "it depends on context";
* No tool guidance — the agent forgets to call the tool and guesses;
* No output format — every response shaped differently, no consistency;
* No constraint — the agent invents numbers when the data is missing.

The next two labs make you write all four. First with one tool, then with three.

**Takeaway:** A persona is a contract. Specify it once, get the same agent every time.

A hands-on course for non-technical managers and operators who need to turn raw business data into decisions, not summaries. You'll learn how to frame analytical questions sharply, read spreadsheets and internal docs through AI, combine internal numbers with market context, and build your own analyst agent that produces decision-ready briefs end-to-end.

Introduction to business analytics through the lens of AI — why raw data doesn't make decisions, the three layers of analytical questions, and where an AI assistant actually fits in the workflow.



Train the framing reflex. Turn vague business questions into sharp ones, pick the metric that actually matters, and lock down time windows, segments, and comparisons before touching any data.



Hands-on labs with the three sources every operator hits: sales spreadsheets, strategy docs in Notion, and market context from the web. Single-tool labs first, then a combined investigation.



Move from running labs to building your own. Write the persona, wire up the tools, and create an agent that delivers weekly KPI briefs and pre-meeting decision packs — finishing with a capstone Quarterly Business Review.



The Persona of an Analyst Agent

A Persona Has Four Parts

What Drifts When You Skip a Part

The Persona of an Analyst Agent

A Persona Has Four Parts

What Drifts When You Skip a Part