File: 'analytics-team-capacity.csv'\nRows: 32 · Columns: week, team_member, hours_logged, active_projects\n\nFull data:\nWeek 1, Maria, 46, 2\nWeek 1, Josh, 44, 2\nWeek 1, Priya, 45, 3\nWeek 1, Tom, 47, 2\nWeek 2, Maria, 48, 2\nWeek 2, Josh, 47, 2\nWeek 2, Priya, 46, 3\nWeek 2, Tom, 50, 3\nWeek 3, Maria, 50, 3\nWeek 3, Josh, 49, 2\nWeek 3, Priya, 48, 3\nWeek 3, Tom, 53, 3\nWeek 4, Maria, 52, 3\nWeek 4, Josh, 51, 3\nWeek 4, Priya, 50, 3\nWeek 4, Tom, 56, 3\nWeek 5, Maria, 54, 3\nWeek 5, Josh, 53, 3\nWeek 5, Priya, 52, 3\nWeek 5, Tom, 59, 3\nWeek 6, Maria, 55, 3\nWeek 6, Josh, 54, 3\nWeek 6, Priya, 53, 3\nWeek 6, Tom, 61, 4\nWeek 7, Maria, 56, 3\nWeek 7, Josh, 55, 3\nWeek 7, Priya, 54, 3\nWeek 7, Tom, 62, 4\nWeek 8, Maria, 57, 3\nWeek 8, Josh, 56, 3\nWeek 8, Priya, 55, 3\nWeek 8, Tom, 63, 4\n\nAuto-detected: team average climbed from 45.5 hrs/week in Week 1 to 57.75 hrs/week in Week 8, with no week showing a decline. Tom has exceeded 60 hours for four straight weeks and is now carrying 4 active projects, the highest sustained load on the team.

Summary  
This chapter covers how to configure a custom AI agent by crafting a system prompt and integrating an external tool so the agent retrieves real data, reasons with it, and formats its response into structured sections.  

General domain of usage  
Hiring decision advisory

You've reasoned through it cold, then grounded it in real capacity data. Package it so the next headcount question doesn't start from scratch.



## What You're Building

An agent that always checks real capacity data before answering, argues the case against its own recommendation, and flags what it doesn't know about budget or timeline.



A hiring decision advisor, not a generic assistant

Always check real capacity data before answering, never guess at workload

Recommendation, case against it, open unknowns

Never state a workload trend or number unless the data actually returned it

Combined into one system prompt:

> <b>You are a hiring decision advisor. When the user asks whether to hire or redistribute work, use the spreadsheet_analyze tool to check the team's actual capacity data before answering — never guess at workload trends from what the user describes alone. Always return exactly three sections: Recommendation (hire, redistribute, or something else, stated plainly), Strongest Case Against It (argue against your own recommendation using what the data actually shows), Open Unknowns (anything about budget, timeline, or the team's specific situation that the data can't tell you). Never state a workload trend, hours figure, or risk unless the data actually returned it.



> ## Task
> - Click "Create Agent."
> - Agent name → 
>    ><b>Hiring Decision Advisor
> - System prompt → paste the block above
> - Choose tools → check Spreadsheet Analyze
> Save, then activate the agent for this chapter
> - Send: 
>    ><b>Should we hire another analyst or redistribute the work? A new hire would run about $95K fully loaded.
>
> Check your result: did it cite Tom's 63 hours and 4 active projects, and the 8-week upward trend, or just repeat generic tradeoffs about headcount cost? 

**Analytics Team Capacity Log**

| Week | Team Member | Hours Logged | Active Projects |
|---|---|---|---|
| Week 1 | Maria | 46 | 2 |
| Week 1 | Josh | 44 | 2 |
| Week 1 | Priya | 45 | 3 |
| Week 1 | Tom | 47 | 2 |
| Week 2 | Maria | 48 | 2 |
| Week 2 | Josh | 47 | 2 |
| Week 2 | Priya | 46 | 3 |
| Week 2 | Tom | 50 | 3 |
| Week 3 | Maria | 50 | 3 |
| Week 3 | Josh | 49 | 2 |
| Week 3 | Priya | 48 | 3 |
| Week 3 | Tom | 53 | 3 |
| Week 4 | Maria | 52 | 3 |
| Week 4 | Josh | 51 | 3 |
| Week 4 | Priya | 50 | 3 |
| Week 4 | Tom | 56 | 3 |
| Week 5 | Maria | 54 | 3 |
| Week 5 | Josh | 53 | 3 |
| Week 5 | Priya | 52 | 3 |
| Week 5 | Tom | 59 | 3 |
| Week 6 | Maria | 55 | 3 |
| Week 6 | Josh | 54 | 3 |
| Week 6 | Priya | 53 | 3 |
| Week 6 | Tom | 61 | 4 |
| Week 7 | Maria | 56 | 3 |
| Week 7 | Josh | 55 | 3 |
| Week 7 | Priya | 54 | 3 |
| Week 7 | Tom | 62 | 4 |
| Week 8 | Maria | 57 | 3 |
| Week 8 | Josh | 56 | 3 |
| Week 8 | Priya | 55 | 3 |
| Week 8 | Tom | 63 | 4 |

Spreadsheet for the Task

Master the skills needed to build reliable AI decision systems with Claude, tackling real business calls like switching tools, approving a hire, and renewing a vendor contract. Progress from prompt-only reasoning to tool-grounded analysis with live data, then package your approach into a reusable Claude agent with structured output and a no-invented-facts constraint. By the end, you'll write a full agent system prompt from scratch.

Discover how to build an AI decision system for a real tool-switch call. Start with prompt-only reasoning, ground it in live pricing and migration data via web search, then build a Tool Switch Advisor agent that routes tasks and cites its sources.

Explore the essentials of contextual prompting for hiring decisions. Compare a bare prompt to a context-rich one, analyze real team capacity data to test the case for a hire, then build a Hiring Decision Advisor agent around that data.

Dive into the fundamentals of building a Claude agent for a vendor renewal decision. Combine past notes, usage data, and market research into one recommendation, then write your own agent prompt with tool routing and a no-invented-facts rule.