Skip to main content
Effective prompting transforms the FinOps Agent from a basic assistant to a powerful FinOps analyst. This guide outlines three core building blocks for designing prompts that create productive and accurate interactions. The difference between vague requests and actionable results comes down to how well you structure your prompt with clear intent, scope, and direction.

1. Intent

Defines what you want the Agent to do; ask questions or perform actions.

2. Scope

Specifies which workspace, provider, data source, timeframe, and groupings to use.

3. Direction

Establishes output format and additional notes for how results are presented.

​
1. Intent

Every productive interaction with the FinOps Agent starts with a clear action or question. The Agent supports two intent categories: ask the Agent questions or ask the Agent to perform actions.

​
Ask the Agent Questions

You can ask questions about your cloud costs or about Vantage features and functionality. 
@Vantage What drove our AWS EC2 costs last month?
@Vantage Which Azure services are trending up this week?
@Vantage Show me any compute-related anomalies for the past week.
@Vantage Can you explain the difference between virtual tags and segments? When should I use one over the other?

You:
@Vantage What drove our AWS EC2 costs last month?

The Agent:
EC2 Cost Drivers — Last Month

Scope used: Production Workspace (token wrkspc_a1b2c3d4e5f6g7h8)
Period: 2024-10-01 to 2024-10-31
Filter: AWS provider, service Amazon Elastic Compute Cloud - Compute
Total EC2 Compute Spend: $127,450.23

Key Takeaways
• One AWS account (123456789012) in us-west-2 drove ~94.2% of EC2 compute costs.
• All EC2 compute spend was in us-west-2.
• Other accounts contributed minimal EC2 compute spend.

Breakdown by Account and Region

| AWS Account ID | Region    | EC2 Compute Cost (USD) | Share of EC2 | Notes                |
|----------------|-----------|------------------------|--------------|----------------------|
| 123456789012   | us-west-2 | $120,058.12            | 94.2%        | Primary cost driver  |
| 987654321098   | us-west-2 | $6,892.45              | 5.4%         | Secondary contributor|
| 555666777888   | us-west-2 | $412.18                | 0.32%        | Minor                |
| 111222333444   | us-west-2 | $87.48                 | 0.07%        | Minor                |
| 999888777666   | us-west-2 | $0.00                  | ~0%          | Negligible           |

What this means
• The vast majority of EC2 compute cost is concentrated in a single account and single region.
• To refine "what drove cost" further (e.g., by usage type/purchase option, or top instances), we can:
  • Group by charge_type to separate On-Demand vs. Savings Plans/RIs and other line items.
  • Group by resource_id to identify the top-cost EC2 instances.
  • Add tag-based groupings (for example, tag:environment) if you use tags for ownership or workload mapping.

Want deeper detail?
• Do you want me to:
  • Break down by purchase option/charge type (On-Demand vs. SP/RI vs. Credits)?
  • List the top-cost EC2 instances in that account?
  • Run the analysis for a different workspace instead of Production?
  • Create a Cost Report you can open in Vantage with this exact filter and date range?

Reply with your preferred next step and, if needed, the workspace name you want analyzed.

​
Ask the Agent to Perform Actions

You can also use the Agent’s tools to generate reports, create Terraform code, or create artifacts. 
@Vantage Create a cost report filtered to AWS Lambda for the last 30 days and put it in the Infra folder
@Vantage I have a list of teams in a CSV file (e.g., integrations, security, etc.). I have an AWS tag called team with value of the same name and a Kubernetes label of organization/team with a value of the same team name. I want to combine all these costs to one virtual tag. Write me a Terraform script to do this.

You:
@Vantage can you create a report for me in the Infra folder (fldr_abc123def4567890) that shows me AWS EC2 box usage for October 2025?

The Agent:
Cost Report Created: AWS EC2 Usage (October 2025)

Your report is ready in the Infra folder and focuses on AWS EC2 instance (compute) usage for October 2025.

• Direct link: https://console.vantage.sh/go/rprt_xyz789abc123def4

Report Summary:
- **Title**: AWS EC2 Usage - October 2025
- **Folder**: Infra (fldr_abc123def4567890)
- **Filter**: (costs.provider = 'aws' AND costs.service = 'Amazon Elastic Compute Cloud - Compute')
- **Date Range**: 2025-10-01 to 2025-10-31
- **Aggregate By**: usage
- **Workspace**: Production (wrkspc_def456ghi789jkl0)

Notes:
• This targets the EC2 compute service: Amazon Elastic Compute Cloud - Compute, which matches "EC2 box usage."
• It excludes EC2-related "other" costs (like EBS, NAT, data transfer). If you want those included, I can broaden the filter.

Want any tweaks?
• Groupings: Add by account_id, region, or instance type (if tagged).
• Compare periods: Keep or remove previous-period comparison.
• Scope: Restrict to specific AWS accounts or regions.

​
2. Scope

The FinOps Agent needs to know which data to use. Vantage customers often have multiple workspaces, multiple cloud providers, and different access levels. Providing scope upfront can help to eliminate back-and-forth questions.

​
Specify a Workspace

The Agent automatically selects a workspace for your queries if you don’t specify one. Understanding this process helps you know what data you’re viewing.
Workspaces in Vantage separate your provider integrations into distinct cost views. Different integrations can be associated with different workspaces, so the workspace you query determines which cost data is available. Learn more about workspaces.
The Agent chooses a workspace in this order:
  1. If you specify one. If you mention a workspace name or token in your prompt, that workspace is used.
  2. Your last used workspace. If you’ve used a workspace before, the Agent remembers and uses that one.
  3. If you only have one workspace. If your account has just one workspace, that’s automatically used.
@Vantage Show me spend for the last 30 days
The Agent provided a high-level response using the default workspace (Infra). For more specific analysis, you’ll need to provide additional scope, such as @Vantage Show me AWS EC2 costs for the last 30 days in the Production workspace.

​
Specify a Folder

When you create Vantage resources, like reports, you can specify which folder to place them in by providing the folder URL or token (found at the end of the URL). 
@Vantage Create a report in https://console.vantage.sh/reporting/fldr_abc123def4567890 showing AWS RDS costs for August 2025

​
Specify Time Windows

Always include time ranges to avoid ambiguity about which period you’re analyzing. 
@Vantage Provide a breakdown of costs by service for Datadog from July 1 to July 31, 2025.

​
Reference Existing Reports

You can reference existing reports by their token or link, and the Agent will analyze them. 
@Vantage in this report rprt_xyz987abc456def789 what was my biggest change in cost from September 2025 to October 2025, what resources contributed to this?

​
3. Direction

Direction helps the Agent shape the final output. Use one of the following output formats to control how results are presented.
3–6 bullet highlights with key numbers. Best for stakeholder updates and quick insights.
@Vantage [your question with intent and scope] Give a summary of findings.
Concise points or steps. Best for action items and next steps.
@Vantage [your question with intent and scope] List top 5 drivers.
Tabular view with columns. Best for grouped spend/usage data.
@Vantage [your question with intent and scope] Show a table grouped by service and region.
Comma-separated rows. Best for spreadsheet import.
@Vantage [your question with intent and scope] Provide CSV for top 10 services.
Side-by-side or delta view. Best for before/after and period-over-period analysis.
@Vantage [your question with intent and scope] Compare last_month vs this_month, highlight deltas.
Markdown checkboxes. Best for runbooks and to-dos.
@Vantage [your question with intent and scope] Output a checklist of fixes.
Highlights with brief rationale. Best for root cause storytelling.
@Vantage [your question with intent and scope] Call out anomalies and likely causes.
VQL, CLI commands. Best for reproducibility.
@Vantage [your question with intent and scope] Show the VQL you used.

​
Putting It All Together

Now that you understand Intent, Scope, and Direction, use this template to combine all three elements into effective prompts. Fill in the brackets with your specific values. 
Intent: [Question | Action]
Scope: workspace=[name or token], folder=[URL or token], provider=[aws|azure|gcp|custom_provider:<token>], services=[...], accounts=[account_id,...], regions=[...], tags=[key=value,...], timeframe=[last_30_days|YYYY-MM-DD..YYYY-MM-DD], groupings=[service,region,account_id,...], date_bin=[day|week|month]
Direction: output=[table|list|chart|summary], notes=[spike threshold, anomaly detection, comparisons]

​
Ready-to-Use Prompts

Copy and customize these prompts for common FinOps tasks. Each prompt follows the Intent, Scope, and Direction framework for optimal results.
Use these prompts to identify what caused unexpected cost increases.
@Vantage What caused the cost spike in the [workspace] workspace [time period]?
Compare actual spending against budgets and forecasts.
@Vantage List all budgets in the [workspace] workspace and compare against actuals
Find unusual spending patterns that may indicate issues.
@Vantage Find anomalies for the last [days] days in the [workspace] workspace
Group and filter costs using tags for allocation and showback.
@Vantage Show cost grouped by tag:[key] for [time period] in [workspace]
Deep dive into existing reports to understand cost changes.
@Vantage In this report [report_token or link], what was my biggest change in cost from [period 1] to [period 2], and what resources contributed to this?

​
Understand Responses

Because the FinOps Agent relies on an LLM, responses are not always deterministic. This means that the same prompt may return slightly different answers, and results heavily depend on the context you’ve provided. Understanding how responses work helps you verify results and get the most out of your interactions with the Agent in cases where you receive unexpected responses.
The FinOps Agent uses Vantage Query Language (VQL) to query your cost data. VQL is a SQL-like language for filtering cloud cost data across providers. When the Agent generates responses, it creates VQL queries behind the scenes to fetch the data.
You can verify Agent responses using the following methods:
  • Compare against Cost Reports: Use your existing Cost Reports as a source of truth. For example, if you ask about EC2 spend last month, compare the answer against your saved report for that timeframe.
  • Request the VQL query: Ask the Agent to show you the VQL it used. You can do this by adding “Show the VQL you used” to your prompt, or by asking in a follow-up message. Review the VQL to understand what data the Agent accessed and confirm the filters and groupings align with your intent.

​
Limitations

Understanding the Agent’s limitations helps you set appropriate expectations and get the most out of your interactions.
The Agent’s external training data goes up to October 2024. It may miss newer changes, such as recent cloud pricing updates or product releases after this date.
The Agent doesn’t browse the internet unless a tool explicitly provides that information. It can’t fetch live pricing from websites unless you provide it or a tool returns it.
The Agent can only access what the Vantage MCP or Vantage Docs MCP tools return and what you share in chat.
The Agent can’t run commands, deploy code, or change infrastructure. It can suggest provider commands, but you must run them yourself.
The Agent may misinterpret ambiguous requests. It might misread a vague error without full context.
The Agent only sees the current chat context provided by the platform. It doesn’t remember previous sessions unless the platform includes them in the current thread.
The Agent cannot analyze images at this time.
The Agent doesn’t access personal data unless you provide it explicitly. It can’t read your emails, tickets, or dashboards without you pasting content or granting access via tools.